<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom" xml:lang="en">
  <title type="text">Andy Fitzgerald Consulting</title>
  <subtitle type="text">Perspectives, methods, notes and presentations about information architecture and structured content design for the modern web.</subtitle>
  <link rel="self" type="application/atom+xml" href="https://preview.andyfitzgeraldconsulting.com/feed.xml"/>
  <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/"/>
  <id>https://preview.andyfitzgeraldconsulting.com/feed.xml</id>
  <updated>2026-11-05T00:00:00Z</updated>
  <author>
    <name>Andy Fitzgerald</name>
  </author>
  <entry>
    <title type="text">Taxonomy Manager Tree View Improvements</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/taxonomy-manager-tree-view-improvements/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/taxonomy-manager-tree-view-improvements/</id>
    <published>2026-11-05T00:00:00Z</published>
    <updated>2026-11-05T00:00:00Z</updated>
    <summary type="text">Usability and management enhancements for iterating and evolving your taxonomies in Sanity Studio</summary>
    <category term="Project Note"/>
    <category term="Taxonomy"/>
    <category term="Sanity"/>
    <content type="html">&lt;p&gt;This week I released a new version of the Taxonomy Manager plugin. It includes usability and management features that should make iterating and evolving your taxonomies in Sanity Studio easier:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Tree view now supports drag and drop hierarchy editing&lt;/li&gt;&lt;li&gt;Top level concepts can be promoted to or demoted from &amp;quot;Top Concept&amp;quot; in the tree view&lt;/li&gt;&lt;li&gt;You can now add existing concepts as well as create new concepts from the tree view&lt;/li&gt;&lt;li&gt;Publication status is now included in the tree view&lt;/li&gt;&lt;li&gt;Concepts open for editing are highlighted in the tree view to help you easily spot where they are in the hierarchy&lt;/li&gt;&lt;li&gt;Content Releases are supported for those on qualifying Enterprise plans&lt;/li&gt;&lt;/ul&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/ed26a2f5d14f58d8405e619161a22b98e1283b33-3024x1898.png?rect=118,228,2906,1670&amp;amp;w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;690&quot; alt=&quot;A multicolumn view of the Sanity Taxonomy Manager plugin showing concept scheme and concept lists, a hierarchy view of the selected taxonomy, and a detail view of the selected term. &quot; /&gt;&lt;figcaption&gt;Sanity Taxonomy Manager Plugin version 6.2.0&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;This update was built with accessibility in mind from the start: the mouse-based interaction is an enhancement of functionality you&amp;#x27;ll find in the overflow menu for each concept, which offers full support for keyboard and screen reader users.&lt;/p&gt;&lt;p&gt;These features also support you freewheeling polyhierarchy enthusiasts out there. Concepts that appear in multiple places in the same hierarchy can be moved independently of each other, and when the concept is open you&amp;#x27;ll see a highlight on both instances (or more ... though good luck with that!)&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/projects/#sanity-taxonomy-manager&quot;&gt;Read more about the Sanity Taxonomy Manager plugin&lt;/a&gt; &lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">Designing and Implementing Semantic Data Layers</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/designing-and-implementing-semantic-data-layers/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/designing-and-implementing-semantic-data-layers/</id>
    <published>2026-10-19T00:00:00Z</published>
    <updated>2026-10-19T00:00:00Z</updated>
    <summary type="text">A practical, actionable, architectural approach to making meaning explicit, consistent and durable.</summary>
    <category term="Book Note"/>
    <category term="Semantic Web"/>
    <category term="Interoperability"/>
    <category term="Technology Integration"/>
    <category term="Data Management"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/3d26476b6785a9f620571be25ef48f8c22735899-334x500.png?bg=e82976&amp;amp;w=952&amp;amp;h=500&amp;amp;fit=fill&amp;amp;auto=format&quot; width=&quot;952&quot; height=&quot;500&quot; alt=&quot;Designing and Implementing Semantic Data Layers book cover&quot; /&gt;&lt;figcaption&gt;&lt;a href=&quot;https://www.goodreads.com/book/show/255520401&quot;&gt;Designing and Implementing Semantic Data Layers&lt;/a&gt;, by Dave Wells. Technics Publications, 2026&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Aligning meaning across systems is a perennial enterprise problem. Tools get added, systems diverge, abstraction layers are “optimized” for pressing use cases, and shortcuts are taken. And that’s before mergers, market pivots, and shifts in strategy. &lt;/p&gt;&lt;p&gt;The result is a growing collection of what different people and systems mean when they say “customer,” “department,” or “protocol.” Dave Wells offers a practical, actionable solution using semantic layers: “a systematic, architectural approach to making meaning explicit, consistent and durable” (121).&lt;/p&gt;&lt;h2 id=&quot;key-concepts&quot;&gt;Key Concepts&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;“Trust in data is not a sentiment. It’s a capability, and semantics is how you build it” (12).&lt;/li&gt;&lt;li&gt;Separation of meaning from structure is crucial for building semantic layers that scale (21). &lt;/li&gt;&lt;li&gt;Definitions are required: “A model without definitions is only a diagram. A model with definitions is a semantic artifact” (46).&lt;/li&gt;&lt;li&gt;The hard work of aligning on definitions is organizational and interpersonal. “Getting subject matter experts to agree on a single definition of &lt;em&gt;customer&lt;/em&gt;” is challenging, but necessary — and can make the difference between an initiative that succeeds and one that stalls (88).&lt;/li&gt;&lt;li&gt;“A semantics-based architecture is not a single semantic layer [but rather a] deliberate network of semantic layers, each serving a specific purpose” (96). &lt;/li&gt;&lt;li&gt;Start small; provide value at each stage: “Incremental delivery is not a compromise. It is the right strategy” (120). &lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;impressions&quot;&gt;Impressions&lt;/h2&gt;&lt;p&gt;This is a brief book (122pp) packed with actionable ideas and strategies. Wells makes a convincing case for why semantic layers are more important now than ever, introduces the core concepts, then gets into the details of five kinds of semantic layers and their respective purposes. He then uses relatable use cases to illustrate how these layers are implemented across typical enterprise data systems. &lt;/p&gt;&lt;p&gt;A running theme is that you don’t create “a” semantic layer; you need an overlapping and interconnected set of layers that are designed and governed to achieve specific outcomes. Wells provides rubrics for choosing which types of layers are best suited for common outcomes. &lt;/p&gt;&lt;p&gt;Wells provides a brief introduction to semantic technology concepts, like RDF triples and ontologies, but if these are new concepts to you a more expansive introduction to the nuts and bolts of knowledge graphs might help you envision the practical applications more concretely. As Wells emphasizes throughout, however, the strategy he describes is “not a technology product. It’s a discipline, practiced through careful modeling, governed through clear ownership, and sustained through organizational commitment” (121). &lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">Site Updates: Structure, Semantics &amp; Standards</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/site-updates-structure-semantics-standards/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/site-updates-structure-semantics-standards/</id>
    <published>2026-10-08T00:00:00Z</published>
    <updated>2026-10-08T00:00:00Z</updated>
    <summary type="text">Four lessons learned from redeveloping my consulting website with LLM code collaboration.</summary>
    <category term="Perspective"/>
    <category term="Content Modeling"/>
    <category term="Structured Content"/>
    <category term="Taxonomy"/>
    <category term="Accessibility Evaluation"/>
    <category term="Performance Evaluation"/>
    <category term="AI Integration"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/543a1b6350fe005a669dc5fe4481ec14ae507655-2400x3200.jpg?rect=0,1460,2400,1350&amp;amp;w=1200&amp;amp;h=675&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;Greyscale photography of concrete blocks&quot; /&gt;&lt;/figure&gt;&lt;p&gt;Last week I cut over to a new version of my consulting website. This iteration includes a new short-form &amp;quot;Notes&amp;quot; content type, a richer approach to publishing presentations, and improvements to topical navigation and search. It also includes structural and design updates that make it the most accessible version of the site I&amp;#x27;ve ever built.&lt;/p&gt;&lt;p&gt;It&amp;#x27;s also the first version for which I entrusted the authoring of my site&amp;#x27;s Typescript almost entirely to an LLM. I defined and articulated decisions about the content model, component boundaries, URLs, semantics, and accessibility, and then allowed the LLM to build, test, and iterate the templates and integration scripts that instantiate those choices in code. As Claude Code framed it, &amp;quot;I owned the &lt;em&gt;seams&lt;/em&gt;, and I delegated the &lt;em&gt;surfaces&lt;/em&gt;.&amp;quot;&lt;/p&gt;&lt;p&gt;As you would expect, this division of labor allowed me to get more done in less time. I can write Typescript, but I&amp;#x27;m abysmally slow at it and tend to get mired in Wagnerian struggles with the type-checker. I also don&amp;#x27;t enjoy it. Delegating that work to an LLM was a gift.&lt;/p&gt;&lt;p&gt;The time savings, however, wasn&amp;#x27;t the real benefit. Delegating &amp;quot;surface&amp;quot; work to an LLM significantly changed the quality of the final output I serve to site visitors. The time I might have otherwise spent debugging scripts or puzzling out server configs was now time I could use to explore navigation models, refine search indexing, interconnect content, and expand my coverage of accessibility and linked data affordances.&lt;/p&gt;&lt;p&gt;Looking over this project in retrospect, I can identify four reasons the &amp;quot;seams and surfaces&amp;quot; division led to a higher quality output than I would have otherwise achieved. Here&amp;#x27;s a brief look at each of them.&lt;/p&gt;&lt;h2 id=&quot;structured-content&quot;&gt;Structured Content&lt;/h2&gt;&lt;p&gt;My site has used structured content, built and served with &lt;a href=&quot;https://www.sanity.io/&quot;&gt;Sanity&lt;/a&gt;, since 2022. Since its first iteration, my content model has been structured around the formal characteristics of each content type, as opposed to the page templates on which content appears. This deliberately decouples the way content is structured from the way it is presented on any given page, which in turn shifts the process of adding new types from one of &amp;quot;rebuilding&amp;quot; the model to one of &amp;quot;recomposing&amp;quot; elements already in place.&lt;/p&gt;&lt;p&gt;As an example, on the previous version of my site I listed all of my past talks by venue and date on a single page. To get details on the topic of a talk, a visitor had to click through to the venue link and read the talk description (assuming the conference site was still live, of course). For the redesign I added descriptions, slide decks and, where available, recordings and transcripts to each presentation. I also gave each presentation a detail page of its own.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/b6ebf4eaf06c59bd2c4e02e0dcc00392961c9210-3020x1884.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;749&quot; alt=&quot;Two versions of the Talks/Presentations page, with the earlier version listing only talk locations and dates, while the updated version includes images, descriptions, decks, and more. &quot; /&gt;&lt;figcaption&gt;Talks listed on the 2024 version of the site and the more detail rich display of Presentations on the 2026 redesign. &lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;While I did have to create a new &amp;quot;Presentation&amp;quot; type to record new information, the &amp;quot;Event&amp;quot; type that was driving the previous view needed only minor modifications to accommodate new fields. Instead of being the source of the listing (as on the old page), this type was recomposed as a dependent part of the new Presentation type.&lt;/p&gt;&lt;p&gt;By defining this new &amp;quot;seam&amp;quot; of the revised content model up front, the LLM building the templates for these new pages already had all the typed data that appears on them, structured to reflect the conceptual shift in how I had remodeled that content. Using typescript as a translation layer from schema design to front-end implementation gave the LLM a machine-readable specification for mapping my Figma design into composable templates (the &amp;quot;surface&amp;quot;).&lt;/p&gt;&lt;p&gt;There was still iteration and fine tuning in the template creation process, but a clear and intentionally structured content schema gave the model far less to guess at or get wrong. This meant I could spend my time reconciling the gaps between content schema and UI design and focus on building a more harmonious system overall.&lt;/p&gt;&lt;h2 id=&quot;semantics&quot;&gt;Semantics&lt;/h2&gt;&lt;p&gt;The formal characteristics of a content type are useful for understanding how to structure and articulate a content model. They do nothing, however, to describe what a given piece of content is about, what it means. That&amp;#x27;s the role of semantics. For the last few years I&amp;#x27;ve managed the semantic relationships between content instances with standards-based taxonomies, built and governed in Sanity using the &lt;a href=&quot;https://github.com/andybywire/sanity-plugin-taxonomy-manager&quot;&gt;Taxonomy Manger plugin&lt;/a&gt; (which I wrote and maintain 😉).&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/92e200a276d1ae845939c5fb9df47d7b742a125e-2996x1858.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;744&quot; alt=&quot;A view of a plugin operating in the Sanity platform, displaying a hierarchical list of terms, and detail page of an individual term.&quot; /&gt;&lt;figcaption&gt;The Taxonomy Manager plugin running in Sanity Studio.&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Managing content semantics and content structure separately allows each of these elements to co-evolve with minimal disruption to the other — and provides a clear seam that helps an LLM not muddle the two.&lt;/p&gt;&lt;p&gt;For example, I&amp;#x27;ve long wanted to include faceted filtering on my &amp;quot;Insights&amp;quot; page. With a seam that clearly distinguished content semantics from content structure, I was able specify the pattern and logic that afford that functionality in Figma, then let the LLM find the best way to code that into a reusable component.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/400924dfbf5104d4f4c0014e98c0525b583d98dd-3020x1884.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;749&quot; alt=&quot;The Insights and Presentations pages from the redesigned site, showing similar facet filter options. &quot; /&gt;&lt;/figure&gt;&lt;p&gt;Since the semantic layer is independent of content types, the pattern worked equally well for the Presentations page, which had also gotten more complex with the redesign.&lt;/p&gt;&lt;p&gt;The separation of content semantics from content structure likewise meant that the seam defined by faceted filtering applied equally well to the Search Results page, which combines all types and genres across both top level nav categories.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/a04b5221d373a6c81d315161b4d248a92116525d-2996x1858.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;744&quot; alt=&quot;The search results page from the redesigned website, showing a similar set of facet filters to the Insights and Presentations pages.&quot; /&gt;&lt;/figure&gt;&lt;p&gt;In each of these cases, the separation of structure from semantics allowed me to more easily integrate the new and revised types into the established system of meaning and create a clear seam to pass the the LLM for collaboration on the surface work.&lt;/p&gt;&lt;h2 id=&quot;standards&quot;&gt;Standards&lt;/h2&gt;&lt;p&gt;The taxonomies described above are structured in alignment with the &lt;a href=&quot;https://en.wikipedia.org/wiki/Simple_Knowledge_Organization_System&quot;&gt;Simple Knowledge Organization System (SKOS)&lt;/a&gt; standard, the W3C recommendation for taxonomies, thesauri, and other kinds of controlled vocabularies. Standards-based vocabularies are interoperable with other semantic web technologies by design, which means that extending them to new types and technologies is accounted for from the start. Since these standards are well-established and open, it also means that LLMs already know how to interpret and build with them.&lt;/p&gt;&lt;p&gt;Established and open standards were likewise an important &amp;quot;seam&amp;quot; in ensring accessibility across the new site. Well-structured, standards-compliant semantic HTML already does most of the heavy lifting of providing accessible content to visitors using assistive technologies. As with SKOS, LLMs already have broad context on implementing standards-compliant HTML and using ARIA attributes to fill in the inevitable gaps that remain.&lt;/p&gt;&lt;p&gt;By specifying a seam for template design that prioritized standard markup, the LLM was able to use the context built into its training corpus to recommend semantic and standards-compliant components, and even pushed back on ideas that could have created accessibility dead zones, keyboard traps, or overly verbose screen reader descriptions.&lt;/p&gt;&lt;p&gt;Compared to the previous site, Lighthouse accessibility scores improved from a range of between 84 and 96 to 100 across all pages, and &lt;a href=&quot;https://wave.webaim.org/&quot;&gt;WAVE&lt;/a&gt; and manual browser testing indicate a broad improvement in structural elements, use of ARIA roles and properties, and clear navigation affordances.&lt;/p&gt;&lt;h2 id=&quot;delegate-and-validate&quot;&gt;Delegate and Validate&lt;/h2&gt;&lt;p&gt;As this point, it&amp;#x27;s fair to ask: &amp;quot;But ... is the generated code any good?&amp;quot; As is common with LLM-driven development, it created more code (nearly 3x more) than I had in the previous version of the site. This iteration does, however, use a different static site generator (Astro versus 11ty, my decision), includes enhanced UI features, live previews, and CI testing for accessibility, link integrity, and performance budgets. Some growth in the codebase isn&amp;#x27;t a huge surprise, but the increase is non-trivial.&lt;/p&gt;&lt;p&gt;The average &lt;em&gt;page&lt;/em&gt; weight the site delivers to site visitors, on the other hand, is less than 20% of the previous load (from 1,817 KB to 222 KB median first load). It is also more accessible than the previous version and uses several hundred megabytes less space on my server (from 630 MB to 94MB).&lt;/p&gt;&lt;p&gt;Could the code be less verbose than it is? Perhaps. More importantly, though, I can account for and explain the purpose of every script and component, and I can map each to a design intent. To me this clarity warrants the extra weight in the repository. It reassures me that I haven&amp;#x27;t (knowingly) included any &amp;quot;black boxes&amp;quot; that may or may not turn out to be bombs later.&lt;/p&gt;&lt;p&gt;There are times I enjoy writing code by hand (not Typescript, of course). I&amp;#x27;ve missed that a bit on this project. But choosing to delegate to the LLM the tasks the LLM is good at has allowed me to spend more time on the tasks that I&amp;#x27;m good at, and that has been surprisingly uplifting. It&amp;#x27;s also helped me build in a way that&amp;#x27;s different than I would have done on my own. Not just faster, but clearer, lighter, and more accessible for the folks who visit my site. It&amp;#x27;s helped me build the kind of website I actually want to use, and of that, I&amp;#x27;m a fan.&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">The Calm Before the Storm</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/the-calm-before-the-storm/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/the-calm-before-the-storm/</id>
    <published>2026-09-19T00:00:00Z</published>
    <updated>2026-09-19T00:00:00Z</updated>
    <summary type="text">Geoff Teehan on the particular moment we’re in with AI assisted software creation: developer scarcity is fading, but we still don&apos;t understand what abundance looks like. </summary>
    <category term="Web Clipping"/>
    <category term="AI Integration"/>
    <category term="Project Management"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/43dc886abb7d72a9ee00fe02a973be8f4c1607ab-480x252.png?rect=32,0,448,252&amp;amp;w=448&amp;amp;h=252&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;448&quot; height=&quot;252&quot; alt=&quot;Margaret Hamilton with the Apollo Guidance Computer listings, 196&quot; /&gt;&lt;figcaption&gt;&lt;a href=&quot;https://design.lightspark.com/we-built-companies-around-the-cost-of-building&quot;&gt;We Built Companies Around the Cost of Building&lt;/a&gt;, lightspark.com&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Geoff Teehan on the particular moment we’re in with AI assisted software creation:&lt;/p&gt;&lt;blockquote&gt;When digital cameras showed up, people used them like film cameras. Same pictures, they just stopped paying to develop them. Nobody was photographing their parking spot yet. That’s about where software is. Faster development is the same pictures without the developing costs. The parking-spot phase hasn’t started.&lt;/blockquote&gt;&lt;p&gt;Are we really prepared for a world where software creation requires the same amount of forethought as taking “fourteen nearly identical photographs of your lunch”? We’re soon to find out!&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">Why Shortcuts Aren&apos;t the Road</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/why-shortcuts-arent-the-road/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/why-shortcuts-arent-the-road/</id>
    <published>2026-09-04T00:00:00Z</published>
    <updated>2026-09-04T00:00:00Z</updated>
    <summary type="text">As with most shortcuts, there’s a reason that the road isn’t laid down that way.</summary>
    <category term="Web Clipping"/>
    <category term="User Research"/>
    <category term="Project Management"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/e217d4862260bb5a3a43094bbf0f45579e925e16-480x282.png?rect=0,6,480,270&amp;amp;w=480&amp;amp;h=270&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;480&quot; height=&quot;270&quot; alt=&quot;Product Picnic label superimposed over a drawing of an orange and an apple. &quot; /&gt;&lt;figcaption&gt;&lt;a href=&quot;https://productpicnic.beehiiv.com/p/taste-is-a-dead-end&quot;&gt;Taste is a dead end; what you should have is a point of view&lt;/a&gt;, productpicnic&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Pavel Samsonov on the perils of “fake research”: &lt;/p&gt;&lt;blockquote&gt;As with most shortcuts, there’s a reason that the road isn’t laid down that way.&lt;/blockquote&gt;&lt;p&gt;Yes, by “fake” he means research an LLM made up. &lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">GraphCon 2026 Themes &amp; Takeaways</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/graphcon-2026-themes-takeaways/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/graphcon-2026-themes-takeaways/</id>
    <published>2026-08-05T00:00:00Z</published>
    <updated>2026-08-05T00:00:00Z</updated>
    <summary type="text">Summary of the top themes and insights from the 2026 GraphCon conference in Seattle.</summary>
    <category term="Conference Themes"/>
    <category term="Taxonomy"/>
    <category term="Knowledge Graphs"/>
    <category term="Technology Integration"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/6395e1c1688e999402714b579635f2de60f5e404-1024x768.jpg?rect=0,192,1024,576&amp;amp;w=1024&amp;amp;h=576&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;1024&quot; height=&quot;576&quot; alt=&quot;A view of a presenter and her title slide from six rows back in an office space turned small conference venue. &quot; /&gt;&lt;/figure&gt;&lt;p&gt;Two weeks ago I had the pleasure of attending the first GraphCon Seattle event. This inaugural practitioners conference, planned and produced by the Graph Geeks community, focused on how folks are solving real world problems in practical ways, using a range of technologies and approaches. Though a one day event, it was packed with thought-provoking perspectives and rich with opportunities to talk with other graph enthusiasts exploring and trying out new ideas with graph approaches to information and knowledge management.&lt;/p&gt;&lt;p&gt;Two weeks ago I had the pleasure of attending the first &lt;a href=&quot;https://www.graphgeeks.org/graphcon&quot;&gt;GraphCon Seattle&lt;/a&gt; event. This inaugural practitioners conference, planned and produced by the &lt;a href=&quot;https://www.graphgeeks.org/&quot;&gt;Graph Geeks&lt;/a&gt; community, focused on how folks are solving real world problems in practical ways, using a range of technologies and approaches. Though a one day event, it was packed with thought-provoking perspectives and rich with opportunities to talk with other graph enthusiasts exploring and trying out new ideas with graph approaches to information and knowledge management.&lt;/p&gt;&lt;p&gt;&lt;/p&gt;&lt;p&gt;Graph Geeks generously plans to make videos of all the talks available on &lt;a href=&quot;https://www.youtube.com/@GraphGeeksOrg&quot;&gt;their YouTube channel&lt;/a&gt;, so keep an eye on that space to check out the talks first-hand. In the meantime, here’s a rundown of the themes and takeaways that emerged for me across the day’s talks and across my conversations with other practitioners.&lt;/p&gt;&lt;h2 id=&quot;top-themes-takeaways&quot;&gt;Top Themes &amp;amp; Takeaways&lt;/h2&gt;&lt;h3 id=&quot;1-value-drives-adoption-no-one-cares-that-youre-using-a-graph&quot;&gt;1. Value drives adoption — no one cares that you’re using a “graph”&lt;/h3&gt;&lt;p&gt;This theme will strike you as ridiculously obvious, because “of course value drives adoption.” We’re all smart, reasonable professionals, making sound decisions based on solid information. Right? Setting aside high irony for a moment, there’s another challenge for graph enthusiasts: we’re &lt;em&gt;really&lt;/em&gt; excited by the technology. This means that even when we know we shouldn’t, it’s hard to resist talking to someone about “the graph.” As Brad Bebee with Amazon Neptune points out, “graph” itself is not an intrinsic benefit. Building support for the solutions graphs provide means finding and clearly articulating the benefits those solutions will provide for the organization — and (ideally) stopping there.&lt;/p&gt;&lt;p&gt;&lt;/p&gt;&lt;p&gt;It’s this last part that is hard for graph practitioners to follow through on. Himanshu Singh recounted that in the process of moving from proof-of-concept to day-to-day graph operations with Netflix, they removed the word “graph” from messaging to stakeholders and collaborators. They even went so far as to remove the word from the name of their project group. Singh advises against “onboarding” your users. You should instead meet them where they are. If people using your solution have to learn a query language or the care and feeding of graph models, you’ve failed. The better outcome is one where they don’t know they’re using graph technology at all. (Mercifully there are events like GraphCon where the graph nerds can properly nerd.)&lt;/p&gt;&lt;h4 id=&quot;talks&quot;&gt;&lt;strong&gt;Talks&lt;/strong&gt;&lt;/h4&gt;&lt;ul&gt;&lt;li&gt;“The Business of Graphs in the Age of AI,” Brad Bebee, AWS Neptune&lt;/li&gt;&lt;li&gt;“The Last Mile of Graphs: Lessons from Production at Netflix,” Himanshu Singh, Netflix&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;2-graphs-are-evolving-systems&quot;&gt;2. Graphs are evolving systems&lt;/h3&gt;&lt;p&gt;This one was kind of obvious … and kind of not. For those of us coming from the IA and taxonomy world, of course models and implementations have to evolve. An information model is no more “one and done” than is quarterly business planning (hence, you know, its quarterly-ness). But there are some aspects of graph work that can feel settled: top-level ontologies and industry standard taxonomies change slowly and provide stability against fast-changing market pressures. In his talk, “Connected Intelligence: How Graphs Transform AI,” David Hughes makes the argument that in intelligently connected systems, AI should also transform the graph. As Hughes puts it, evolving systems should “return the consequences [of an operation] back to memory.” These “remembered consequences” become part of the graph.&lt;/p&gt;&lt;p&gt;&lt;/p&gt;&lt;p&gt;Yohei Nakajima refers to this as “moving toward an ever-evolving ontology.” In this scenario, ontologies are never “finished,” but are rather a way to represent and understand the state of an agent. The benefit is to mitigate or eliminate the limitations of session-based AI in a scalable and transparent way. Dave Bechberger provides some of my favorite descriptions of this problem: as Bechberger puts it, AI’s “goldfish problem” stems from “session amnesia” and a “re-derivation tax,” where discovered facts must be continually rediscovered. Markdown context documents are a workaround for session-to-session memory, but it’s with graphs that we can approach durable memory for long running processes. Bechberger again: “Vectors remember what. Graphs remember how it relates, and when it was true.”&lt;/p&gt;&lt;h4 id=&quot;talks&quot;&gt;&lt;strong&gt;Talks&lt;/strong&gt;&lt;/h4&gt;&lt;ul&gt;&lt;li&gt;“Connected Intelligence: How Graphs Transform AI,” David Hughes, HDC&lt;/li&gt;&lt;li&gt;“Graphs After RAG: From Knowledge to Context,” Yohei Nakejima, Untapped Capital&lt;/li&gt;&lt;li&gt;“From Goldfish to Elephant: Using Graphs for Long Term Memory,” Dave Bechberger, Stitch Fix&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;3-its-all-about-the-models&quot;&gt;3. It’s all about the models&lt;/h3&gt;&lt;p&gt;One of my favorite motifs across talks was the idea that the real value graph-minded work delivers is not a set of technologies, but rather the models that graph technologies are used to represent. Brad Bebee makes the point that “you don’t need a graph database to use a graph model.” Indeed, the collection of Markdown documents many folks are using for session-to-session AI context are essentially graph models already. Bebee advises that the opportunity before us as graph practitioners is to “be ready to help [these folks] do this better.”&lt;/p&gt;&lt;p&gt;&lt;/p&gt;&lt;p&gt;As an IA/UX professional who came into the field when mobile browsers were just starting to gain traction, this combination of model centrality and learning systems maps cleanly in my mind to a practice our industry has honed over many years: progressive enhancement. This is the idea that the systems we design, be they responsive websites or agentic workflows, should &lt;em&gt;start&lt;/em&gt; with the basic, core functionality we intend to deliver, they layer on speed, performance, and aesthetic &lt;em&gt;enhancements&lt;/em&gt; as the context of use allows. Crucially, should any of these enhancements fail, the core, basic functionality must remain operational. When combined with AI, a graph is an &lt;em&gt;enhancement &lt;/em&gt;of the underlying model. Building these models right, and ensuring that they evolve consistently over time, is far more important than which database we choose or whether we represent data as triples or nodes, edges, and properties.&lt;/p&gt;&lt;h4 id=&quot;talks&quot;&gt;&lt;strong&gt;Talks&lt;/strong&gt;&lt;/h4&gt;&lt;ul&gt;&lt;li&gt;“The Business of Graphs in the Age of AI,” Brad Bebee, AWS Neptune&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;4-polyglot-systems&quot;&gt;4. Polyglot systems&lt;/h3&gt;&lt;p&gt;There’s an ongoing (and certainly faux) controversy in some corners of the graph community over which representation of graph data is “better”: Resource Description Frameworks (RDF) or Labeled Property Graphs (LPG). I was delighted to discover that the GraphCon community was having none of this “controversy” nonsense. Himanshu Singh described a project his team ran on a semantic (RDF) graph on the back end, transformed to a property (LPG) graph on the front end, and which was then accessed by end-users that didn’t know they were using a “graph” at all. David Bechberger likewise frames the roles of vector stores and graph stores as fundamentally complementary: each supports the other by filling in where the other is weak — and getting out of the way when it’s poorly suited for a given task.&lt;/p&gt;&lt;p&gt;&lt;/p&gt;&lt;p&gt;As someone who learned a second language as an adult, I have a personal appreciation for the benefits of polyglotism. Having a second language at one’s disposal opens up new worlds of literature, film, and culture — to say nothing of one’s experience as a traveler. Beyond these first order benefits, however, the &lt;em&gt;process&lt;/em&gt; of learning and operating in a new language generates new insight into one’s native language. Once I had to learn to speak and think outside of English, my understanding of what makes it work (and what makes it fail) changed entirely.&lt;/p&gt;&lt;p&gt;&lt;/p&gt;&lt;p&gt;When applied to how we model data, a polyglot approach likewise offers promise for both first and second order benefits. Moving between, for example, RDF and LPG, allows us to choose the right set of capabilities for the task at hand. It also encourages us to scrutinize the underlying model that each is meant to represent and challenge the decisions and structures that result from thinking primarily in one idiom or the other.&lt;/p&gt;&lt;h4 id=&quot;talks&quot;&gt;&lt;strong&gt;Talks&lt;/strong&gt;&lt;/h4&gt;&lt;ul&gt;&lt;li&gt;“The Last Mile of Graphs: Lessons from Production at Netflix,” Himanshu Singh, Netflix&lt;/li&gt;&lt;li&gt;“From Goldfish to Elephant: Using Graphs for Long Term Memory,” Dave Bechberger, Stitch Fix&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;5-informed-delegation&quot;&gt;5. Informed delegation&lt;/h3&gt;&lt;p&gt;As an event focused on practitioners and real world solutions, GraphCon presentations did not shy away from code examples and live demos. Talks and hallway conversations likewise grappled head-on with the question of how much code one actually writes any more, as opposed to generating with the assistance of LLMs. BoundaryML CEO Vaibhav Gupta stated the case for machine generation directly: “it’s no longer possible to keep up with code subsystems … but we can constrain them.” Indeed, as I understand it, that’s the whole purpose of BAML, the “agent-first” language his company creates.&lt;/p&gt;&lt;p&gt;&lt;/p&gt;&lt;p&gt;Hallway conversations with developers likewise revealed a surprisingly uniform acquiescence to the changing role of individuals in the code creation process. One developer I spoke with said he hardly writes anything by hand any more — and doesn’t miss it. Instead he prefers spending more time with the problem solving and strategy work that informs the scaffolding and constraints he puts in place for the LLM. We’re still in the early days of sorting out what these agent assisted workflows look like — and how to make them consistently reliable — but one thing I didn’t run into were devs pining away for the good old days of writing error handling code and hunting down missing semicolons.&lt;/p&gt;&lt;h4 id=&quot;talks&quot;&gt;&lt;strong&gt;Talks&lt;/strong&gt;&lt;/h4&gt;&lt;ul&gt;&lt;li&gt;“Code is Just a Graph,” Vaibhav Gupta, Boundary (BAML)&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;the-year-of-the-graph&quot;&gt;The year of the graph?&lt;/h2&gt;&lt;p&gt;There was an ongoing joke throughout the day about how many consecutive years over the last few decades have been proclaimed “The Year of the Graph” … and about how flatly each of those predictions have fallen. This was a laid-back enough crowd that I felt like everyone enjoyed the self-deprecating humor. At the same time, there was a palpable sense that graph approaches and technologies are finally finding a compelling, widespread use case as a complement to generative LLMs in the way that modern software is created and maintained. Is this finally “the year of the graph”? I won’t venture a guess, but if it is there’s a good chance that those doing the work won’t use the word “graph” in describing the value they deliver.&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">Structured Content in Product Design</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/presentations/structured-content-in-product-design/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/presentations/structured-content-in-product-design/</id>
    <published>2026-04-09T00:00:00Z</published>
    <updated>2026-04-09T00:00:00Z</updated>
    <summary type="text">Learn concepts and frameworks for understanding structured content design as a systems practice</summary>
    <category term="Workshop"/>
    <category term="Structured Content"/>
    <category term="Mental Models"/>
    <category term="Content Reuse"/>
    <category term="Governance"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/bf004ed2e11bfe498c56db49f95d303b0a2606e2-1920x1140.jpg?rect=0,30,1920,1080&amp;amp;w=1200&amp;amp;h=675&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;Button Workshop Structured Content in Product Design, facilitated by Andy Fitzgerald&quot; /&gt;&lt;/figure&gt;&lt;p&gt;This workshop will introduce concepts and frameworks for understanding structured content design as a systems practice, and will build on those concepts with hands-on exercises based on real-world examples of content rich applications and websites. The approaches we&amp;#x27;ll cover apply to content-rich environments including product interfaces, CMS-driven websites, and emerging AI-powered experiences like knowledge bases. The concepts are platform-agnostic and not tied to any specific technology.&lt;/p&gt;&lt;h2 id=&quot;workshop-highlights&quot;&gt;Workshop Highlights&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;Key principles of structured and composable content design.&lt;/li&gt;&lt;li&gt;Frameworks for structured content design and change management in content rich environments.&lt;/li&gt;&lt;li&gt;Technical and implementation considerations for structured content delivery.&lt;/li&gt;&lt;li&gt;Collaboration strategies for working with engineering, design, and product stakeholders.&lt;/li&gt;&lt;/ul&gt;</content>
  </entry>
  <entry>
    <title type="text">AI Assisted Content Operations in Practice</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/presentations/ai-assisted-content-operations-in-practice/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/presentations/ai-assisted-content-operations-in-practice/</id>
    <published>2026-03-19T00:00:00Z</published>
    <updated>2026-03-19T00:00:00Z</updated>
    <summary type="text">A chat with Sanity builders on the systems, AI workflows, and experiments they&apos;ve built with Sanity. </summary>
    <category term="Panel"/>
    <category term="Sanity"/>
    <category term="AI Integration"/>
    <category term="Structured Content"/>
    <content type="html">&lt;p&gt;In this Sanity Pioneers fireside chat from SF, Meaza (Engineering Manager at Airtable, previously at Figma) and Andy Fitzgerald (Content Operations Consultant at Elemeno Health/WHO) share the systems, AI workflows, and experiments they&amp;#x27;ve built with Sanity. &lt;/p&gt;&lt;h2 id=&quot;panel-highlights&quot;&gt;Panel Highlights&lt;/h2&gt;&lt;p&gt;Conversation topics include:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Testing LLMs against professional taxonomists&lt;/li&gt;&lt;li&gt;Experimenting with Skills for &amp;quot;non-human-shaped problems&amp;quot; &lt;/li&gt;&lt;li&gt;Getting whole teams to think in schemas&lt;/li&gt;&lt;li&gt;Bulk AI translation with human review gates&lt;/li&gt;&lt;li&gt;Migrating content with MCP in weeks instead of months.&lt;/li&gt;&lt;/ul&gt;&lt;h2&gt;Watch the Panel&lt;/h2&gt;&lt;p&gt;&lt;a href=&quot;https://www.youtube.com/watch?v=1-8awGw1mIc&quot;&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/fbf929c09cf640239bb323aeaffc65efcf91a81c-1280x720.jpg?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;Andy, Meaza, and Evelina in a panel conversation&quot; /&gt;&lt;/a&gt;&lt;/p&gt;&lt;p&gt;YouTube · 29:19&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">SHACL for the Practitioner</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/shacl-for-the-practitioner/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/shacl-for-the-practitioner/</id>
    <published>2025-11-15T00:00:00Z</published>
    <updated>2025-11-15T00:00:00Z</updated>
    <summary type="text">A non-academic introduction to SHACL with rich examples and real-world real world case studies. </summary>
    <category term="Book Note"/>
    <category term="Knowledge Graphs"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/a7c9fad0b6caf074b7a9c704ece994b317e409ef-284x400.png?bg=c1c0c2&amp;amp;w=761&amp;amp;h=400&amp;amp;fit=fill&amp;amp;auto=format&quot; width=&quot;761&quot; height=&quot;400&quot; alt=&quot;SHACL for the Practitioner front cover&quot; /&gt;&lt;figcaption&gt;&lt;a href=&quot;https://www.goodreads.com/book/show/245855210-shacl-for-the-practitioner&quot;&gt;SHACL for the Practitioner&lt;/a&gt;, by Veronika Heimsbakk. Self-published, 2025&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;SHACL is the World Wide Web Consortium’s (W3C) standard for validating semantic web data. Standards like RDFS and OWL use inference to discover “new” information in graph data based on logical implications and “open world” assumptions. SHACL allows practitioners to constrain the conclusions semantic inference can draw based on established rules for graph entities and attributes. &lt;/p&gt;&lt;h2 id=&quot;key-concepts&quot;&gt;Key Concepts&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;SHACL can be used to catch logical inferences that are “correct” according to your semantic model, but make no sense in the real world.&lt;/li&gt;&lt;li&gt;SHACL is complementary to OWL, not a replacement for it (35). &lt;/li&gt;&lt;li&gt;SHACL doesn’t change your graph. When run, the SHACL engine delivers a report detailing constraint violations (76).&lt;/li&gt;&lt;li&gt;Most commercial semantic web tools support core SHACL constraints, though their implementation of the full standard still varies. &lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;impressions&quot;&gt;Impressions&lt;/h2&gt;&lt;p&gt;My enthusiasm for this (admittedly niche) topic is twofold: First, &lt;em&gt;SHACL for the Practitioner&lt;/em&gt; means that we now have a complete set of approachable, reader-friendly books for the semantic web technologies that form the foundations of knowledge graphs (RDFS, OWL, SPARQL, SHACL).&lt;br/&gt;&lt;br/&gt;Second, Veronika&amp;#x27;s clear, accessible style takes what could quickly become a technically obtuse topic and frames it as a pragmatic, down to earth complement to OWL and RDFS.&lt;br/&gt;&lt;br/&gt;While SHACL isn&amp;#x27;t otherwise a starting place I would recommend to someone new to semantic web technologies, part one of this book, &amp;quot;Back to Basics&amp;quot;, is hands down the clearest holistic overview I&amp;#x27;ve seen of what the semantic stack seeks to accomplish and how the technologies fit together. Even if you need to come back to the rest of the book for a second pass after you&amp;#x27;ve covered the other technologies elsewhere, this initial framing provides an admirable map for the rest of the semantic web territory.&lt;br/&gt;&lt;br/&gt;Parts two and three provide an introduction to the nuts and bolts of designing, writing, and applying SHACL to validate data (of course), and then a selection of real-world case studies (&amp;quot;SHACL Stories&amp;quot;) that illustrate the wide variety of applications for which SHACL provides practical data validation at scale. The coverage of SHACL shapes and constraints is concise and to-the-point, and kept relatable with a common example throughout. Some of the case studies were more accessible to me than others, but I suspect this is just right: everyone will bring their own background and experience to these stories, and there&amp;#x27;s a range of topics here to accommodate that.&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">Sanity Everything*[NYC] Conference Themes &amp; Takeaways</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/sanity-everything-nyc-conference-themes-takeaways/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/sanity-everything-nyc-conference-themes-takeaways/</id>
    <published>2025-09-30T00:00:00Z</published>
    <updated>2025-09-30T00:00:00Z</updated>
    <summary type="text">Conference and courtyard themes from Sanity’s first “Everything*[NYC]” developer conference.</summary>
    <category term="Conference Themes"/>
    <category term="Sanity"/>
    <category term="Technology Integration"/>
    <category term="AI Integration"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/b4a27c48ff3f1bb4548590b4815cca20abd978af-4032x3024.png?rect=321,722,2925,1645&amp;amp;w=1200&amp;amp;h=675&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;Conference MC Simeon Griggs giving a behind the scenes look at the Sanity Learn platform&quot; /&gt;&lt;/figure&gt;&lt;p&gt;Last week I had the pleasure of attending &lt;a href=&quot;https://sanity.io/&quot;&gt;Sanity&lt;/a&gt;’s first annual “Everything*[NYC]” developer conference. This one-day event was packed with perspectives, feature demos, and customer stories showcasing the leaps forward Sanity has made in the last year—and shining a light down the path of where they’re headed in the rapidly evolving content creation and management space.&lt;/p&gt;&lt;p&gt;Everything*[NYC] was a &lt;em&gt;product &lt;/em&gt;conference, as opposed to a discipline or industry conference. The sessions and demos were thus expectedly rosy. I’m a fan of both the tools Sanity creates and the philosophy that informs them, so this all landed well with me. The consulting work I do, however, focuses first and foremost on my clients’ needs, so I was likewise keen to explore the edges of the programmed narrative with Sanity developers, customers, potential customers, and fellow agency partners in courtyard conversations throughout the day.&lt;/p&gt;&lt;p&gt;In that spirit, read on for my themes and takeaways from the event. I’ll cover what I’m seeing (and am enthused about) from Sanity, what I heard from peers and colleagues in the courtyard, and share a nerdy milestone that I’m counting as a small personal victory in my quest to wean folks away from the idea that “content” necessarily means “pages.”&lt;/p&gt;&lt;h2 id=&quot;platform-themes&quot;&gt;Platform Themes&lt;/h2&gt;&lt;p&gt;The single overarching theme for the day as a whole was “AI” (i.e. LLM) integration into the Sanity platform. Sanity co-founders Magnus Hillestad and Simen Svale et this context with a morning keynote focused on the difference between &lt;em&gt;replacing &lt;/em&gt;work with technology (i.e. most approaches in the recent “AI” deluge) and &lt;em&gt;rewiring &lt;/em&gt;work with technology. They drew a parallel to the (first) industrial revolution, where technology initially displaced individual workers, then, in a fast-follow second act, changed the infrastructure of how work was done altogether, creating new economies of scale and new opportunities.&lt;/p&gt;&lt;h3 id=&quot;agent-actions&quot;&gt;Agent Actions&lt;/h3&gt;&lt;p&gt;Agent Actions provide one example of how Sanity is building infrastructure to not just use AI, but to rewire content work around it. Agent Actions allow developers to code functions that use integrated LLMs to perform defined operations on content using other content as context. This last piece is what piques my interest most: prompts can be managed as documents, and actions can programmatically search and use other content in a collection as context.&lt;/p&gt;&lt;p&gt;As an example, The Met shared an agent action they use to automatically evaluate content against their ten page style guide as part of the publication process. One &lt;em&gt;could &lt;/em&gt;do this with NLP, or one &lt;em&gt;could &lt;/em&gt;try to get authors to apply these rules assiduously, but in my experience, the former approach tends to be fussy and brittle, and the latter approach makes me feel like a schoolmarm with a penchant for hitting children with rulers. (You’re welcome for that mental image.) The agent action means that these checks happen for every document that gets published, and it does so without requiring additional review steps from content producers.&lt;/p&gt;&lt;p&gt;A second case study, presented by &lt;a href=&quot;https://www.linkedin.com/in/laurenherdman/&quot;&gt;Lauren Herdman&lt;/a&gt; of Mejuri, demonstrated how Mejuri uses an agent action to systematically find, reuse, and repurpose content across their collection to increase consistency and cut down on duplication. At first glance this might sound routine, but the struggle of organizations to understand what content they have and avoid recreating content that already exists is real—and costly. As with The Met example above, these are checks that can be applied every time a piece of content is changed. This is work that is theoretically possible with the right discipline and structure … but which I’ve never seen pulled off at scale in the real world.&lt;/p&gt;&lt;h3 id=&quot;reshaping-content-problems-to-fit-the-new-wiring&quot;&gt;Reshaping Content Problems to Fit the New Wiring&lt;/h3&gt;&lt;p&gt;Everything*[NYC] was explicitly billed as a developer conference, so I wasn’t too surprised to see the content design decisions that make the new tech work getting less stage time. Sean Grove (OpenAI &amp;amp; Netlify) did, however, discuss the need to “reshape the problems we address” to take advantage of emerging AI infrastructure by destructuring them to run across multiple agents. Grove’s thesis focused on the development process and creating actionable specifications for AI agents, but this sentiment rings equally true to the way we structure information in the schemas and content models with which those agents operate.&lt;/p&gt;&lt;p&gt;In my experience, many organizations—including those using cutting edge platforms like Sanity—still tend to think of their content as “pages.” The Met style guide example from above, clocking in at “ten pages,” is a case in point. If that guide is like any other I’ve seen, it is composed of an extensive set of directives that work together systematically to create an intended outcome. Loading that ten page document into an LLM context window is certainly a first step in adjusting to this new infrastructure, but it also brings the problem’s old “shape” to our shiny new solution. Sanity has given developers the ability to model information like this as classified, distinct assertions from day one. There’s a bit more planning and set-up involved, but this provides a shape that can be indexed, managed, and composed at scale, both for humans (yes, often as pages) and as data for the bots increasingly among us.&lt;/p&gt;&lt;p&gt;This brings me to a corollary to the “rail” example of the industrial revolution. Railway right-of-ways &lt;em&gt;literally &lt;/em&gt;paved the way for the uninhibited place-to-place connections needed for telegraph, which had repercussions not possibly imagined by those laying the first rail lines. Will integration with AI provide the right-of-way for destructuring our content into systematically related propositions, as opposed to the blobby page construct that’s been with us since moveable type nudged us into mass literacy? We’ll see!&lt;/p&gt;&lt;h2 id=&quot;courtyard-themes&quot;&gt;Courtyard Themes&lt;/h2&gt;&lt;p&gt;Across the several conversations I had with current and prospective Sanity customers and fellow agency colleagues, two themes emerged with conspicuous regularity: folks are still struggling with the chaos of connecting content across product silos, and folks are still locked deep in a mindset of “content as pages.”&lt;/p&gt;&lt;p&gt;As with the writing guide example discussed above, the tools for breaking out of the “pages” mindset are already at hand. Bridging silos is a pricklier task (see &lt;a href=&quot;https://www.linkedin.com/in/davemccomb/&quot;&gt;Dave McComb&lt;/a&gt;’s &lt;em&gt;Software Wasteland &lt;/em&gt;and&lt;em&gt;The Data-Centric Revolution &lt;/em&gt;for all the gory details), but shares a similar root cause with the “pages” problem: idiosyncratic information models across across silos (and pages) make connecting the facts, concepts, and ideas they share in common a manual task.&lt;/p&gt;&lt;p&gt;A large enough language model could maybe make some of those connections with almost reasonable accuracy, but those same models can also help us structure content in a way that simpler models (or simply taxonomies) could make those connections with far greater accuracy.&lt;/p&gt;&lt;p&gt;As above: We’ll see! I’m encouraged, however, that, even if some organizations’ inertia keeps them from following an “ideal” path toward structured content, we at least have the tools for taking incremental steps toward creating the structural and semantic foundations that can keep our newly wired AI workplaces from going completely, well, “off the rails.” (Sooo many metaphors 🙄)&lt;/p&gt;&lt;h2 id=&quot;one-small-nerd-victory&quot;&gt;One Small Nerd Victory&lt;/h2&gt;&lt;p&gt;This is only a “theme” inasmuch as it occurred more than once, but occasionally, when I introduced myself to someone new, they would respond with, “Oh, you’re the taxonomy guy.” Guilty as charged! 🎉&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">Ranked Lists of Related Content in Sanity Studio</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/presentations/ranked-lists-of-related-content-in-sanity-studio/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/presentations/ranked-lists-of-related-content-in-sanity-studio/</id>
    <published>2025-09-15T00:00:00Z</published>
    <updated>2025-09-15T00:00:00Z</updated>
    <summary type="text">A hands-on demo of approaches for generating ranked lists of related content using taxonomy tags in Sanity Studio</summary>
    <category term="Talk"/>
    <category term="Taxonomy"/>
    <category term="Content Discovery"/>
    <category term="Sanity"/>
    <content type="html">&lt;p&gt;A hands-on demo of three simple approaches I use to generate ranked lists of related content using taxonomy tags in Sanity Studio.&lt;/p&gt;&lt;h2 id=&quot;talk-highlights&quot;&gt;Talk Highlights&lt;/h2&gt;&lt;p&gt;See examples of ways to rank related content by:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Tag Count: How many taxonomy tags a related resource shares with a parent resource&lt;/li&gt;&lt;li&gt;Relatedness: What percentage of total tags a related resources shares with a parent resource&lt;/li&gt;&lt;li&gt;Curation: Applying curated boost or suppression values to specific classes of resources to meet the content coordination needs of a particular campaign or initiative&lt;/li&gt;&lt;/ul&gt;&lt;h2&gt;Watch the Talk&lt;/h2&gt;&lt;p&gt;&lt;a href=&quot;https://www.youtube.com/watch?v=-A8FfdfkFRw&amp;amp;list=PLh6eGbGKdvNJFJezFP-eo7uoj7ZI2Phb8&quot;&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/a3f12cb45997fee7d5770752605aa491013141f9-1280x720.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;Screenshot of Andy speaking with the list tag count, relatedness, and curation to his left&quot; /&gt;&lt;/a&gt;&lt;/p&gt;&lt;p&gt;YouTube · 5:10&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">Content Modeling in the Age of Flying Cars</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/content-modeling-in-the-age-of-flying-cars/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/content-modeling-in-the-age-of-flying-cars/</id>
    <published>2025-08-26T00:00:00Z</published>
    <updated>2025-08-26T00:00:00Z</updated>
    <summary type="text">Structuring and connecting content for the next generation of AI-integrated CMS.</summary>
    <category term="Perspective"/>
    <category term="Content Modeling"/>
    <category term="Structured Content"/>
    <category term="Interoperability"/>
    <category term="AI Integration"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/3f0c188e3734580c4a00c0ac66e7e3b8bae246a2-4096x2304.jpg?rect=468,183,3159,1777&amp;amp;w=1200&amp;amp;h=675&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;Three paper airplanes suspended by threads against a cerulean blue background.&quot; /&gt;&lt;/figure&gt;&lt;p&gt;Next-generation content management platforms use generative AI to help content teams work beneath and beyond the traditional “page” — but only if you can reliably tell them how your content is connected in meaningful, repeatable ways. Fortunately, you already have the tools to build the model your new platform needs.&lt;/p&gt;&lt;p&gt;In these heady and sometimes head-spinning few years since GPT3 appeared, those of us responsible for planning, producing, and maintaining content have had a lot to figure out. How do we ensure our content is seen, understood, and trusted now that the internet is being flooded with AI generated output? How do we get trustworthy information to audiences who prefer to use AI mediated tools like Perplexity or “AI Overviews”? How do we effectively and responsibly use LLMs in strategy and production to create higher quality content more efficiently?&lt;/p&gt;&lt;p&gt;In response to these and many other questions, content management system (CMS) vendors have been busy expanding their features and functionality. We now have access to built-in AI assisted idea generators and ghostwriter assistants. We have integrated machine learning based translation and personalization support for audience segments. Several tools also now have support for integrated MCP servers that connect directly to AI assistants and platform native vector stores that enable semantic search.&lt;/p&gt;&lt;p&gt;For the few vendors that are truly integrating these tools throughout their stack, these aren’t just incremental improvements: they’re pushing leading platforms into an entirely new category. Where traditional CMSes cater to the two dimensional flat plane of the page—what’s on the page, in what order, and what comes before and after—emerging platforms move authors beyond the familiar trappings of headings and layouts to represent facts, concepts, and ideas as richly connected entities in a defined domain of knowledge.&lt;/p&gt;&lt;p&gt;If one were to imagine the traditional, page-based CMS as the serviceable Ford Fiesta that got us reliably from print to online content, the emerging CMS of 2025 is the flying car of content management. These are systems free from the earthbound constraints of the flat page—with all the opportunity and all risk that that entails.&lt;/p&gt;&lt;h2 id=&quot;beyond-and-beneath-the-page&quot;&gt;Beyond and Beneath the Page&lt;/h2&gt;&lt;p&gt;The flying car CMS is not just a collection of bolted-on features. Platforms truly engineered for AI integration allow producers and strategists to work confidently in two new directions on an entirely new axis: beyond the page and beneath the page.&lt;/p&gt;&lt;p&gt;Features like personalized audience segments, automated translation, entity linking, and flexible serialization (for example to JSON-LD, llms.txt, or agents.md) are examples of working &lt;em&gt;beyond&lt;/em&gt; the page. In many cases it still makes sense to produce pages as well, but working beyond that page means you can represent your content as richly interconnected data files, markdown, or speakable text, all with the same first class support that is traditionally given to pages.&lt;/p&gt;&lt;p&gt;The flying car CMS also enables you to reach &lt;em&gt;beneath&lt;/em&gt; the page, capturing and providing granular access to the facts, concepts, and ideas that come together to form a how-to article, a product description, or an orchestrated multimedia experience. These deeply integrated AI tools give content producers production, tagging, and research superpowers—and free them from the tedious copying and pasting between fields that plagues so many existing systems.&lt;/p&gt;&lt;h2 id=&quot;still-stuck-in-traffic&quot;&gt;Still Stuck in Traffic&lt;/h2&gt;&lt;p&gt;For all the truly stunning advances we’re seeing with modern content platforms, there is one trap I see organizations fall into again and again when implementing a next generation CMS. They’ve bought the new software, they’ve integrated it with their front end, they’ve onboarded their team … but their content model includes the same constraints—and often the same structure—as their previous platform. They’ve loaded everyone into their shiny new flying car, but they’re still navigating with a paper map of roads and highways, stuck in the same gridlock they swore they were leaving behind.&lt;/p&gt;&lt;p&gt;How does this happen? The content model “map” is only part of the problem. Even if you’ve evaluated your content model and created a new, revised content schema in your new platform, if you’ve used the same &lt;em&gt;cartographic elements&lt;/em&gt; you used before—the same trappings and habits of page-based design—you remain limited to the two dimensions the page affords.&lt;/p&gt;&lt;p&gt;If this distinction between schemas and models, or maps and cartography, sounds fussy and pedantic, let me offer two more comparisons to help you relate: Only a truly reckless pilot would take off not knowing the approach vectors, safe altitudes, and navigation aids necessary to safely land. Only a foolhardy skipper would set sail without knowing the depths, underwater hazards, or prevailing currents along their route. A roadmap has affordances for neither of these navigation contexts; aeronautical and marine charts do and afford safe passage for those who use them wisely.&lt;/p&gt;&lt;p&gt;A better map, which is to say a more expressive and robust schema, begins with a more robust and expressive conceptual model. These models are achieved by stepping back from the page and identifying the facts, concepts, and ideas your pages are meant to communicate, and then modeling those elements to support the coordinated ensemble of people, process, and technology (including AI) you use to create content. This process is called Domain Driven Content Modeling.&lt;/p&gt;&lt;h2 id=&quot;domain-driven-content-modeling&quot;&gt;Domain Driven Content Modeling&lt;/h2&gt;&lt;p&gt;Every organization operates within a specific sphere of influence and activity. This is &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/domain-modeling/&quot;&gt;your organization’s “domain&lt;/a&gt;.” By aligning on and articulating the goals you’re working to achieve within your domain, you can more easily identify the relevant attributes and relationships common to your content. These are the facts, concepts, and ideas germane to the actual work you do, the “differences that make a difference” for your organization.&lt;/p&gt;&lt;p&gt;Every organization also produces content with a specific combination of people, process, and technology. This is your approach to “content operations.” Your content operations may include a team of 20 people using established workflows across a highly polished tech stack, or it may be one person periodically updating markdown files on a geriatric Thinkpad. If your organization creates content, you already have an approach to content operations: it&amp;#x27;s either working toward your goals, or it&amp;#x27;s working against them.&lt;/p&gt;&lt;p&gt;Your content model sits at the center of your content operations. It describes your organization’s content structure (types, objects, and attributes) and content semantics (the formal relationships between structural elements, taxonomy terms, and business rules). In a domain driven approach to content modeling, the design of the content model is explicitly informed by the goals your organization is working to achieve within your domain, and the opportunities and constraints inherent to your people, process, and technology. It’s an approach to content modeling that isn’t based on the formal characteristics of the pages you publish, but rather on the &lt;em&gt;work you actually do.&lt;/em&gt;&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/e4b021011e1db06d0c880f22258f4cea006ed98a-2600x1600.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;738&quot; alt=&quot;A Venn diagram with People, Process, and Technology labels and their overlapping parts combined with the label &amp;quot;Content Modeling,&amp;quot; surrounded by a circle labeled &amp;quot;Content Ops,&amp;quot; which is inside a circle labeled &amp;quot;Business Goals,&amp;quot; which is inside a circle labeled &amp;quot;Domain.&amp;quot;&quot; /&gt;&lt;/figure&gt;&lt;p&gt;If you create documentation for medical equipment, for example, a domain driven content model might capture set-up, operation, and troubleshooting steps for a particular machine. These are not stored as blocks of free text under generic headings, but as structured objects in a machine readable sequence associated with defined use cases and equipment models. Likewise, if you produce footwear from a range of different materials, your content model may capture provenance and cleaning information for each type of material. Descriptions for products using these materials can then draw from a central store of facts to describe where the materials were sourced and how to care for them.&lt;/p&gt;&lt;p&gt;Your content producers may still be &lt;em&gt;composing&lt;/em&gt; in a page-based UI—it’s what authors know and are comfortable with—but when the underlying model is not limited by page constraints, CMS integrated AI assistants can draw on the underlying structure and relationships &lt;em&gt;beneath&lt;/em&gt; the page to make informed suggestions, reduce time searching for information, and allow producers to focus on the activities where they produce the most value. Once produced, this content can appear both as a page and &lt;em&gt;beyond&lt;/em&gt; that page as machine readable markup or plain text optimized for generative AI.&lt;/p&gt;&lt;h2 id=&quot;how-to-get-started&quot;&gt;How to Get Started&lt;/h2&gt;&lt;p&gt;As you can by now surmise, the first step in creating a domain driven content model has nothing to do with code. The CMS is often where the domain model is realized as a concrete entity—even in a “flying car” CMS—but this doesn’t mean its creation is an engineering task alone. In addition to meeting technical requirements, the content model should support the needs and capabilities of content producers and coordinate the requirements of your strategy, production, authorization, and publication process. Your organization’s goals within its domain drive the selection of what gets represented and at what level of granularity, and how those elements relate to other parts of the model.&lt;/p&gt;&lt;p&gt;Fortunately, most of the techniques for gathering this information are well established and widely available. Stakeholder alignment, process modeling, contextual inquiry, and content auditing can all provide valuable insight into the cartographic elements needed to structure a well connected, domain driven content model. Methodologies like the &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/structured-content-design-22/&quot;&gt;Structured Content Design Workflow&lt;/a&gt; can help you plan and stage these activities and work toward a domain-based model that allows you to work on, beneath, and beyond the page.&lt;/p&gt;&lt;p&gt;You also don’t need to tackle the whole model at once. The leading vendors in the “flying car” CMS space have built in affordances and workflows for change management over time. Use these to introduce a domain-driven approach in stages—and to support continued evolution once you’ve effectively moved beyond the two dimensional constraints of the page.&lt;/p&gt;&lt;p&gt;Whichever approach you take, consider not only the model you’re using to orchestrate your content operations at present, but also the concepts and “cartographic language” you make available to your model-building process. When it comes to managing content, the long-awaited age of flying cars is (metaphorically) finally here. We’re all still learning how to navigate these new horizons, but one thing that’s certain is that this new territory will need a new kind of map—and a fresh perspective on how we plan, build, and share the content we create.&lt;/p&gt;&lt;p&gt;&lt;br/&gt;&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">Content Operations Strategy and Design</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/content-operations-strategy-design/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/content-operations-strategy-design/</id>
    <published>2025-07-07T00:00:00Z</published>
    <updated>2025-07-07T00:00:00Z</updated>
    <summary type="text">Designing and building systems for planning, creating, publishing, and managing content.</summary>
    <category term="Perspective"/>
    <category term="Content Modeling"/>
    <category term="Governance"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/132a0c0d6760d6634fbfb028723442fa63055ce4-2400x1600.jpg?rect=0,250,2400,1350&amp;amp;w=1200&amp;amp;h=675&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;Road curving through a forest of tall trees in a dark forest with a beam of light shining down&quot; /&gt;&lt;/figure&gt;&lt;p&gt;In the last several years, we’ve witnessed seismic shifts in how content is created, communicated, and consumed. To communicate effectively online, we now need models that transcend the page and treat content as a living asset, ready to flow wherever it creates value. While Information architecture remains essential, the holistic operationalization of how content is planned, created, published, and managed is now more crucial than ever. This is why I’m narrowing the scope of my practice to “content operations engineering and design,” with a focus on work with the Sanity platform.&lt;/p&gt;&lt;p&gt;For the last fifteen years I’ve framed my work primarily as “information architecture”: I help make complex information spaces easy to understand and pleasant to use. This work has always been about communication—helping teams convey facts, concepts, and ideas so audiences understand what matters. Often, the focus has been on creating pages (despite my &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/presentations/conversations-with-robots-voice-ai-structured-content/&quot;&gt;perennial efforts&lt;/a&gt; to move beyond that metaphor).&lt;/p&gt;&lt;p&gt;In the last several years, we’ve witnessed seismic shifts in how content is created, communicated, and consumed. That we’ve moved past “pages” is a foregone conclusion. To communicate effectively online we now need models that treat content as a living asset, ready to flow wherever it creates value. Information architecture remains essential. The holistic operationalization of how content is planned, created, published, and managed, however, is now more crucial than ever. This is why I’m narrowing how I define the scope of my practice to “content operations strategy and design,” with a focus on modern content operations platforms.&lt;/p&gt;&lt;p&gt;I realize this is a bit of a mouthful, so let me unpack it piece by piece.&lt;/p&gt;&lt;h2 id=&quot;content-operations&quot;&gt;Content Operations&lt;/h2&gt;&lt;p&gt;&lt;em&gt;Content operations&lt;/em&gt; is the coordinated ensemble of &lt;strong&gt;people, process, and technology&lt;/strong&gt; an organization uses to plan, create, govern, and measure the effectiveness of content. Purposeful, outcomes-focused content operations guide both human authors and generative algorithms in the creation of trustworthy, effective content.&lt;/p&gt;&lt;p&gt;In my practice, I’ve found that creating a successful approach to content operations relies on three interdependent activities:&lt;/p&gt;&lt;ol&gt;&lt;li&gt;&lt;strong&gt;Discovery and alignment:&lt;/strong&gt; uncovering an organization’s communication goals, domain language, and author realities&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Thinking in systems&lt;/strong&gt;: designing models, services, and workflows that scale beyond individual documents and consider communication goals across contexts&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Engineering for intended outcomes:&lt;/strong&gt; building structure and guidance into the content operations environment so that the creation of on-brand, on-purpose content is the path of least resistance for both humans and machines&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;If you follow my work at all, you know that this is the kind of work I’ve already been doing for years. By formalizing this focus, my goal is to more directly connect with organizations that need help uncovering what must be said, modeling it so it can move fluidly across channels, and embedding those models into workflows that create consistent value.&lt;/p&gt;&lt;h2 id=&quot;strategy&quot;&gt;Strategy &lt;/h2&gt;&lt;p&gt;As &lt;a href=&quot;https://en.wikipedia.org/wiki/Attention_economy#History&quot;&gt;Herbert Simon&lt;/a&gt; observed more than fifty years ago, in an information-rich world, “a wealth of information creates a poverty of attention.” While it has never been easier to create &lt;em&gt;more&lt;/em&gt; content, identifying effective approaches for creating and articulating the &lt;em&gt;right&lt;/em&gt; content reliably and efficiently across channels and contexts remains a challenge. If anything, the proliferation of low quality AI generated content now makes that task even more difficult. &lt;/p&gt;&lt;p&gt;This is where strategy comes in. Creating, publishing, and managing content that furthers organizational goals doesn’t happen on its own. Stakeholders, SMEs, designers, and engineers must be aligned on what those goals are and what constitutes a “win” in achieving them. Teams that are all pulling in different directions can spend a lot of time and energy staying in same spot.&lt;/p&gt;&lt;p&gt;Successful content operations projects build consensus on what they will achieve prior to deciding how they will achieve it. I work with my clients to vet and apply time-tested frameworks for ensuring their strategic approach to content is grounded in measurable business needs, not the latest trends of UI design and front-end development. &lt;/p&gt;&lt;h2 id=&quot;design-engineering&quot;&gt;Design &amp;amp; Engineering&lt;/h2&gt;&lt;p&gt;&lt;a href=&quot;https://articles.centercentre.com/design_rendering_intent/&quot;&gt;Jared Spool&lt;/a&gt; writes that &amp;quot;design is the rendering of intent.” I like to extend this truism by adding that engineering is the realization of design. To move from strategic intent to operational reality, both design and engineering must be carefully coordinated. When misalignment exists between these elements for content operations, organizations can end up with “technically correct” implementations that ignore business requirements and author needs, or visionary design recommendations that never survive handoff.&lt;/p&gt;&lt;p&gt;For &lt;a href=&quot;https://uxdesign.cc/prototyping-information-architecture-11aaccacf9a1&quot;&gt;nearly a decade&lt;/a&gt; I’ve been developing and iterating methods for bridging engineering and design. As part of my focus on content operations, I’ve brought together a set of time-tested frameworks and methods for uncovering and instantiating the requirements for an organization’s content operating systems. These tools allow me to lead projects in which:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Alignment workshops provide informed input for domain and content models.&lt;/li&gt;&lt;li&gt;Domain and content models directly inform the schemas and components of a working content platform.&lt;/li&gt;&lt;li&gt;Content platforms are embedded alongside training and governance guidance that helps teams take full ownership of their content operating system and its evolution.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;The systems that emerge are ones that authors actually want to use because they speak their language, support their tasks, and provide effective tools for embracing new ways of creating and managing content.&lt;/p&gt;&lt;h2 id=&quot;modern-content-operations-platforms&quot;&gt;Modern Content Operations Platforms&lt;/h2&gt;&lt;p&gt;For the past five years most of my client work has been deep inside the Sanity ecosystem: I’ve launched studios, written plugins, developed apps, and designed and iterated custom workflows and integrations. I’ve also had a front-row seat as Sanity and other platforms have evolved from capable headless CMSes into full-featured “content operating systems.” For organizations with specialized entities or domain‑specific complexity, the customizable schemas and API‑first workflows of modern tools make them uniquely suited for creating purpose-built content operations solutions.&lt;/p&gt;&lt;p&gt;The openness and flexibility that makes these systems powerful, however, can backfire: Development teams new to composable content are just as free to create implementations that work &lt;em&gt;against&lt;/em&gt; core value propositions as they are to build in a way that embraces them. By committing to understanding the platform and its features deeply, my goal is to help organizations using Sanity and similar tools work &lt;em&gt;with the grain&lt;/em&gt; of these platforms to achieve the operational outcomes that best serve their needs and deliver lasting value over time.&lt;/p&gt;&lt;h2 id=&quot;carrying-on&quot;&gt;Carrying On&lt;/h2&gt;&lt;p&gt;Twenty years ago I pivoted from a career path as an English professor to the then budding field of “UX.” My goal then, as now, has been to understand and enable the creation and sharing of meaning between individuals. The changes wrought by generative AI in the last few years have tossed a lot of what we thought we knew about those practices into an epistemological juice mixer. The outcomes I seek to support, however, remain the same: the communication of ideas from one mind to another.&lt;/p&gt;&lt;p&gt;I’m excited to carry on with this work in this new, weird world of ours. If any of what I’ve written piques your curiosity—or if you find yourself in the thick of any of the challenges I describe above—please do &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/contact/&quot;&gt;reach out&lt;/a&gt;. I’d love to start a conversation.&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">Orchestrating the Mundane: Deskilling Language Models </title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/presentations/orchestrating-the-mundane-deskilling-language-models/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/presentations/orchestrating-the-mundane-deskilling-language-models/</id>
    <published>2025-05-02T00:00:00Z</published>
    <updated>2025-05-02T00:00:00Z</updated>
    <summary type="text">Learn how to orchestrate language models for consistent, predictable help with workaday IA tasks</summary>
    <category term="Talk"/>
    <category term="AI Integration"/>
    <category term="Automation"/>
    <category term="Technology Integration"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/20c13142c908ca94f3826d41e9d48e8d34f11ae2-1080x1080.png?rect=0,0,1080,608&amp;amp;w=1080&amp;amp;h=608&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;1080&quot; height=&quot;608&quot; alt=&quot;Session card for Andy Fitzgerald at the IA Conference&quot; /&gt;&lt;/figure&gt;&lt;p&gt;What does it take to harness the capabilities of LLMs to help with workaday IA tasks like classification, text analysis, or audits? If you’ve ever tried to do these via a chatbot, you’ve likely already discovered that conversational interfaces don’t scale—and often aren’t very reliable.&lt;/p&gt;&lt;p&gt;Large language models can be harnessed for these tasks, and this session will show you how by combining two techniques: deskilling and orchestration. These approaches help us set up interconnected flows of small, well-defined tasks that are then combined to accomplish more complex outcomes.&lt;/p&gt;&lt;h2 id=&quot;talk-highlights&quot;&gt;Talk Highlights&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;An introduction to the technique of orchestrating language models&lt;/li&gt;&lt;li&gt;A working example of model orchestration in practice&lt;/li&gt;&lt;li&gt;Tactics for advocating for these approaches with colleagues and collaborator&lt;/li&gt;&lt;/ul&gt;</content>
  </entry>
  <entry>
    <title type="text">Taxonomy Management and use in Sanity Studio</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/presentations/taxonomy-management-and-use-in-sanity-studio/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/presentations/taxonomy-management-and-use-in-sanity-studio/</id>
    <published>2025-04-04T00:00:00Z</published>
    <updated>2025-04-04T00:00:00Z</updated>
    <summary type="text">A hands-on demo of how to get started using standards based vocabulary control with Taxonomy Manager in Sanity Studio</summary>
    <category term="Talk"/>
    <category term="Taxonomy"/>
    <category term="Content Modeling"/>
    <content type="html">&lt;p&gt;Are you still managing your &lt;a href=&quot;https://www.linkedin.com/company/sanity-io/&quot;&gt;Sanity&lt;/a&gt; content with free tags, flat lists, or hard-coded values? This 20 min &amp;quot;lightning&amp;quot; talk can help you get started using standards based vocabulary control with Taxonomy Manager in Sanity Studio.&lt;/p&gt;&lt;h2 id=&quot;talk-highlights&quot;&gt;Talk Highlights&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;A concise definition of taxonomy and what it accomplishes&lt;/li&gt;&lt;li&gt;An overview of the 5 most common taxonomy use cases&lt;/li&gt;&lt;li&gt;Why and how to decouple content structure from content semantics&lt;/li&gt;&lt;li&gt;Installing and configuring the Taxonomy Manager plugin&lt;/li&gt;&lt;li&gt;Adding terms and structure to a taxonomy in Sanity Studio&lt;/li&gt;&lt;li&gt;Adding a taxonomy field to a Sanity document &amp;amp; tagging content&lt;/li&gt;&lt;li&gt;Querying content collections based on taxonomy terms&lt;/li&gt;&lt;li&gt;Rules-based inferencing in Studio using taxonomy relationship&lt;/li&gt;&lt;/ul&gt;&lt;h2&gt;Watch the Talk&lt;/h2&gt;&lt;p&gt;&lt;a href=&quot;https://www.youtube.com/watch?v=kkFRFtFSNeQ&quot;&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/2b919a882a69765b8127154bed0e2b1a512a1213-1280x720.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;Sanity Showcase Taxonomy Management and Use in Sanity Studio video poster&quot; /&gt;&lt;/a&gt;&lt;/p&gt;&lt;p&gt;YouTube · 24:21&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">Designing Content Authoring Experiences</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/designing-content-authoring-experiences/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/designing-content-authoring-experiences/</id>
    <published>2025-03-15T00:00:00Z</published>
    <updated>2025-03-15T00:00:00Z</updated>
    <summary type="text">Well designed authoring experiences make doing the right things easy and doing the wrong things hard.</summary>
    <category term="Book Note"/>
    <category term="Governance"/>
    <category term="Structured Content"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/daf74a7c6ebd55a9e8078d07113c9f1fe60682f0-324x500.png?bg=e9eaea&amp;amp;w=952&amp;amp;h=500&amp;amp;fit=fill&amp;amp;auto=format&quot; width=&quot;952&quot; height=&quot;500&quot; alt=&quot;Front cover of Designing Content Authoring Experiences&quot; /&gt;&lt;figcaption&gt;&lt;a href=&quot;https://www.goodreads.com/book/show/223598420-designing-content-authoring-experiences&quot;&gt;Designing Content Authoring Experiences&lt;/a&gt;, by Greg Dunlap. Splash Hit Publishing, 2025&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;&lt;em&gt;Designing Content Authoring Experiences&lt;/em&gt; centers on learning who your users are and how they operate, then designing workflows that make it easy for them to create quality, well-structured content (and that make it difficult to create inchoate jumbles of wordmess). Crucially, Greg Dunlap also provides examples aplenty of what this looks like in practice and methods for how designers can work with authors and stakeholders to achieve these ends in the design process.&lt;/p&gt;&lt;h2 id=&quot;key-concepts&quot;&gt;Key Concepts&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;Good author experience design requires a careful balance of structure and flexibility (19).&lt;/li&gt;&lt;li&gt;“Unless designed very carefully, page builders and block-based editors remove the content’s semantic meaning” (27).&lt;/li&gt;&lt;li&gt;“Removing flexibility is &lt;em&gt;always &lt;/em&gt;harder than adding it. […] Authors will always ask for more” (33).&lt;/li&gt;&lt;li&gt;Authoring tools should focus on making “doing the right things easy and doing the wrong things hard” (45).&lt;/li&gt;&lt;li&gt;A CMS is not a writing tool; it’s a publishing tool (128).&lt;/li&gt;&lt;li&gt;Don’t forget to design training! (131)&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;impressions&quot;&gt;Impressions&lt;/h2&gt;&lt;p&gt;The book&amp;#x27;s examples focus on CMS design, but the insights they provide aren’t limited to those contexts. CMS use cases are legion and there is clearly a lot designers can do to make those experiences better. This book isn&amp;#x27;t about the technical details of any one system, however, and advice for adapting established patterns to user needs focuses on the user needs, not the patterns themselves. It&amp;#x27;s hard to say how CMS design and functionality will evolve over the next few years, but the need for human-friendly ways to interact with those systems will undoubtedly remain.&lt;br/&gt;&lt;br/&gt;Dunlap and his editors have done the hard work of making this book short, engaging, and actionable. In the spirit of the A Book Apart (RIP) publications, this is a brief book for people who make things that can be read over a weekend (or in one sitting if you&amp;#x27;re a fast reader—I&amp;#x27;m not). Having read a few ponderous and wandering tomes in my time, I am grateful for the effort that goes into creating short work that is direct and to the point.&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">AI Supported Auto-Categorization</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/ai-supported-auto-categorization/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/ai-supported-auto-categorization/</id>
    <published>2025-03-03T00:00:00Z</published>
    <updated>2025-03-03T00:00:00Z</updated>
    <summary type="text">A comparison of manual, NLP, and LLM-based tagging methods and key takeaways for effective tagging approaches.</summary>
    <category term="Method"/>
    <category term="Development"/>
    <category term="Automation"/>
    <category term="AI Integration"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/f4f8c74f76f015b1cd5afa5fc32d84843c09bdb4-5184x3456.jpg?rect=410,233,4506,2535&amp;amp;w=1200&amp;amp;h=675&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;Four paper tags of different colors with strings attached on a plain white background&quot; /&gt;&lt;/figure&gt;&lt;p&gt;Generative AI offers some enticing shortcuts for tagging large content sets. But even outputs that &amp;quot;look&amp;quot; right might not meet user goals—and may not be any more effective than what can be generated with time-tested natural language processing approaches. This article describes an experiment that compares these two methods to discover the essential ingredients for an effective tagging strategy.&lt;/p&gt;&lt;p&gt;For the last few weeks, I&amp;#x27;ve been exploring ways that AI and LLMs do—and don&amp;#x27;t—fit into taxonomy design and development for content-driven websites. I&amp;#x27;ve written about:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/ai-taxonomy-term-generation/&quot;&gt;AI Supported Taxonomy Term Generation&lt;/a&gt;, for which I&amp;#x27;ve found there are some compelling use cases for integrating LLMs&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/purpose-driven-taxonomy-design/&quot;&gt;Purpose-Driven Taxonomy Design&lt;/a&gt;, in which the primary challenge I see clients struggle with is defining what the taxonomy needs to accomplish (for which you&amp;#x27;ll need to—gasp!—actually talk to other humans)&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/sanity-studio-taxonomy-migration/&quot;&gt;Migrating a Taxonomy to the CMS&lt;/a&gt;, which can be significantly streamlined when your work is based on established semantic web standards&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Once your taxonomy is integrated into your CMS, the next step, naturally, is to tag your content. This is what allows you to put the domain knowledge represented by your taxonomy to use to recommend content, display collections, navigate, support search, inform analytics, surface content opportunities, etc.&lt;/p&gt;&lt;p&gt;If you already have a collection of several thousand accurately tagged articles and you need to tag more, using an enterprise taxonomy tool to train a machine learning model to tag new content may be a feasible way to get this done. If you&amp;#x27;re not in the fortunate position of having a critical mass of accurately tagged content, your options have traditionally been manually slogging through the content set with human taggers, or using deterministic natural language processing (NLP) algorithms and models to assign tags.&lt;/p&gt;&lt;p&gt;But we live in 2025! Surely LLMs can help with this. Right? If you&amp;#x27;ve been on The Internet lately, you&amp;#x27;ve certainly seen examples aplenty of LLMs creating reasonable-looking sets of tags for content, seemingly out of thin air. But are these tags effective? And are they any better than what a human tagger or NLP process would produce?&lt;/p&gt;&lt;h2 id=&quot;an-experiment&quot;&gt;An Experiment&lt;/h2&gt;&lt;p&gt;Since the content set I&amp;#x27;ve been using as a running example in the last few articles is small enough for one person to tag manually, I took the opportunity to conduct an experiment. I ran human, NLP, and LLM tagging approaches on a sample set of articles to evaluate each approach for consistency.&lt;/p&gt;&lt;p&gt;My hypothesis, based similar experiences I&amp;#x27;ve had with recent client work, was that the (deterministic) NLP approach would be at least as accurate as the LLM approach, and would, in the final implementation, involve less overhead that keeping an LLM from going off the rails.&lt;/p&gt;&lt;p&gt;It turns out that I was wrong.&lt;/p&gt;&lt;p&gt;But I was not wrong in the way I expected to be: A comparison of methods showed not the inherent differences of one algorithmic approach over the other—or even over manual human tagging. Rather, each approach produced a result that &lt;em&gt;looked&lt;/em&gt; right ... but that may or may not be effective for achieving the user-centered goals intended.&lt;/p&gt;&lt;p&gt;In short, I was reminded that you can&amp;#x27;t do human-centered design without input from the humans for whom the design is intended. 🤷‍♂️&lt;/p&gt;&lt;p&gt;In the spirit of sharing this rediscovery of the obvious—and with the suspicion that it might be a useful reminder for others in these heady AI times—here&amp;#x27;s a look at:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;the small sample size experiment I ran with human, deterministic, and LLM-based auto-tagging&lt;/li&gt;&lt;li&gt;the outcomes and the findings produced &lt;/li&gt;&lt;li&gt;a perspective on managing and setting expectations for tagging projects in our nascent era of ubiquitous LLMs&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;auto-tagging-approaches&quot;&gt;Auto-Tagging Approaches&lt;/h2&gt;&lt;p&gt;To establish a baseline for what an appropriate set of tags should include, I first manually tagged all the articles in the content collection in question with four terms each from the new taxonomy. This took roughly four hours (not counting breaks—of which I needed plenty).&lt;/p&gt;&lt;p&gt;I also enlisted two experienced information architecture colleagues to tag the three most difficult and the three easiest articles. They were given the same parameters I used to tag content myself: aim for four tags per article focused on the article&amp;#x27;s overall topic, its &amp;quot;aboutness,&amp;quot; in light of the intended audience and use cases for the content and the site. (My colleagues were much faster than I was—it turns out I&amp;#x27;m a lousy scanner).&lt;/p&gt;&lt;h3 id=&quot;natural-language-processing-auto-tagging&quot;&gt;Natural Language Processing Auto-Tagging&lt;/h3&gt;&lt;p&gt;Natural Language Processing (NLP) uses linguistic models and rules-based interpretation methods to deterministically generate information about natural language texts. The NLP process I set up to tag the six sample articles used &lt;a href=&quot;https://www.sbert.net/&quot;&gt;BERT sentence transformers&lt;/a&gt; to calculate semantic similarity between a given article and the definition of each taxonomy term at three levels:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;Document similarity&lt;/strong&gt; compared the entire article to each taxonomy term definition&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Paragraph similarity&lt;/strong&gt; compared each paragraph to taxonomy term definitions and selected the highest-scoring terms&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Sentence similarity&lt;/strong&gt; compared each sentence to taxonomy term definitions and selected the highest-scoring terms&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;I then weighted these scores for each term, with the document score carrying the most weight (60%) and the maximum sentence score carrying the least (10%). This produced a final similarity score for each term definition, of which the top four were selected as tags. This process took 2.5 hours to set up and run.&lt;/p&gt;&lt;p&gt;👉 &lt;a href=&quot;https://github.com/andybywire/auto-cat/blob/main/nlp-tagging.ipynb&quot;&gt;View the annotated NLP Auto-Tagging Jupyter Notebook&lt;/a&gt;&lt;/p&gt;&lt;h3 id=&quot;large-language-model-auto-tagging&quot;&gt;Large Language Model Auto-Tagging&lt;/h3&gt;&lt;p&gt;Finally, I used OpenAI&amp;#x27;s GPT-4o-mini to tag articles via LLM. Since 4o&amp;#x27;s context window can handle 128,000 tokens, I was able to feed it an entire article &lt;em&gt;and&lt;/em&gt; the entire taxonomy, including definitions, for each article to be tagged. I set the temperature of the model to zero to limit fanciful embellishment and loquacious asides.&lt;/p&gt;&lt;p&gt;As you would expect, it took some experimentation and iteration to get the prompt right, but I was eventually able to get the model to return a set of appropriate tag labels as a JSON file I could compare with the other results. This process took about two hours overall and incurred a total of $0.93 cents in API fees.&lt;/p&gt;&lt;p&gt;👉 &lt;a href=&quot;https://github.com/andybywire/auto-cat/blob/main/llm_tagging.ipynb&quot;&gt;View the annotated LLM Auto-Tagging Jupyter Notebook&lt;/a&gt;&lt;/p&gt;&lt;h2 id=&quot;results&quot;&gt;Results&lt;/h2&gt;&lt;p&gt;To assess the consistency of each approach, I calculated the Jaccard index across all possible combinations of tag results. For two sets of values, which in this case amounts to up to eight tags, the Jaccard index calculates the similarity and diversity between the sets by dividing the intersection of the sets (the count of which values they have in common) by the union of the sets (the total number of unique values between them).&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/d5ab989196ac1a33803f05d684767f31725aee64-4800x1828.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;457&quot; alt=&quot;Two Venn diagrams, one showing the intersection sets, the other showing the union of sets. Between them is a division symbol, indicating that the intersection is divided by the union.&quot; /&gt;&lt;figcaption&gt;A Jaccard index divides the intersection of sets by the union of sets&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;The results, shown in the heat map below, indicate the average degree of agreement between individual tagging agents: three humans, one NLP process, and one LLM process. One is complete agreement. Zero means that the set of tags provide by each were completely different.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/d21986eebea7b5506ce81182fb82ed892889ef0e-1920x1440.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;900&quot; alt=&quot;A Tagger Agreement heatmap, showing that all taggers, both human and algorithmic, agreed no more than 35% and no less than 17%.&quot; /&gt;&lt;/figure&gt;&lt;p&gt;None of these are what I would consider strong results for &lt;em&gt;any&lt;/em&gt; pair—including pairs of human taggers. Likewise, neither the NLP nor the LLM process represent a significant difference in the consistency offered by manual tagging.&lt;/p&gt;&lt;p&gt;So how do we know which set of tags is the most effective for the goals established for this content collection (and outlined in the &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/purpose-driven-taxonomy-design/&quot;&gt;Purpose Driven Taxonomy Design&lt;/a&gt; article)? We would need to know which approach is able, from the user&amp;#x27;s point of view, to:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;coordinate content recommendations&lt;/li&gt;&lt;li&gt;communicate the range of topics covered&lt;/li&gt;&lt;li&gt;afford topic-based browsing&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Even if the three human taggers had stronger agreement on which tags to choose for each article, this &lt;em&gt;still&lt;/em&gt; doesn&amp;#x27;t tell us that those tags will achieve these goals for users in the context in which they&amp;#x27;ll encounter them. To discover this answer, we need to engage not with some new tool or learning model, but rather with the actual users themselves.&lt;/p&gt;&lt;p&gt;As a fairly seasoned UX professional, this finding wasn&amp;#x27;t exactly a shocker. What did take me by surprise was how &lt;em&gt;equally worthless&lt;/em&gt; each of the three techniques I tried above is in the absence of this user insight. With so much attention focused on what LLMs can do, it&amp;#x27;s somehow refreshing to be starkly reminded that human-centered design still requires human involvement in a fundamental way.&lt;/p&gt;&lt;h2 id=&quot;takeaways&quot;&gt;Takeaways&lt;/h2&gt;&lt;p&gt;Fortunately, there are several common and well-established methods for collecting this missing insight.&lt;/p&gt;&lt;h3 id=&quot;prior-to-tagging&quot;&gt;Prior to Tagging&lt;/h3&gt;&lt;p&gt;If it&amp;#x27;s important that your taxonomy and tagging strategy meet improvement goals at launch, you&amp;#x27;ll want to engage with users before you start tagging content, and ideally before you finalize your taxonomy. Two common and well-established methods for accomplishing this are:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Tree testing, which uses a raw, text-based &amp;quot;tree&amp;quot; of topic tags to test out candidate structures with potential users&lt;/li&gt;&lt;li&gt;Prototype usability testing, which presents potential users with a mocked-up version of a content collection (and its hierarchy) and leads them to complete a set of defined tasks with the proposed design&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Both of these are effective, low-cost ways to vet a proposed topic hierarchy and tagging approach prior to implementation. They can also be effective ways to find out if your labels and categories deliver the &amp;quot;information scent&amp;quot; needed to guide users to related topics and content.&lt;/p&gt;&lt;p&gt;These techniques only work, of course, if you actually do them. Asking a bot or a &amp;quot;synthetic user&amp;quot; (🙄) to complete a task has become infinitely easier in recent months, but unless those bots &lt;em&gt;are your users&lt;/em&gt;, you&amp;#x27;re only adding a new, equally ambiguous column to the &amp;quot;Tagger Agreement&amp;quot; heat map above.&lt;/p&gt;&lt;h3 id=&quot;after-tagging&quot;&gt;After Tagging&lt;/h3&gt;&lt;p&gt;For some projects, getting the taxonomy and tagging strategy right at launch is not mission-critical. The content collection I&amp;#x27;m using as an example in these articles falls into this category. In this case, I&amp;#x27;ll use data from the traffic and click path analytics and user feedback to help me collect the insight I&amp;#x27;ll need to iterate and improve the taxonomy and tagging strategy over time.&lt;/p&gt;&lt;p&gt;Indeed, even if you have front-loaded user research with tree testing and usability testing, monitoring a taxonomy&amp;#x27;s performance and adapting it to new uses and contexts is an ongoing pursuit. This process is called governance, and it is indispensable to maintaining a taxonomy&amp;#x27;s effectiveness over time.&lt;/p&gt;&lt;p&gt;As with &amp;quot;before tagging&amp;quot; approaches, evidence-based iteration and purposeful governance only work if you actually &lt;em&gt;do&lt;/em&gt; them. The care and feeding of a taxonomy is a requirement that catches many organizations new to vocabulary control by surprise. It remains, however, vitally important to ensuring the lasting value of taxonomy and tagging efforts.&lt;/p&gt;&lt;h2 id=&quot;setting-expectations&quot;&gt;Setting Expectations&lt;/h2&gt;&lt;p&gt;It would seem, then, that in the absence of front-loaded user research, using an LLM to tag content offers no appreciable advantage over deterministic NLP approaches to auto-categorization. Without user input, neither approach is much more than an exercise in theoretical tag generation.&lt;/p&gt;&lt;p&gt;Since taxonomies are ultimately systems intended to support human behavior in all of its variance and unpredictability, it is hard to imagine an algorithmic approach replacing human creators and maintainers outright. While there certainly appear to be opportunities for automated processes to take some of the tedium and repetition out of this work, it is only by purposefully incorporating the fundamentals of user-centered design that we will achieve results that both &amp;quot;look right&amp;quot; and deliver the outcomes intended for the people who must use them.&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">Sanity Studio Taxonomy Migration</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/sanity-studio-taxonomy-migration/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/sanity-studio-taxonomy-migration/</id>
    <published>2025-02-11T00:00:00Z</published>
    <updated>2025-02-11T00:00:00Z</updated>
    <summary type="text">Instructions, templates, and scripts for migrating a taxonomy from a spreadsheet into the Taxonomy Manager tool on Sanity CMS.</summary>
    <category term="Method"/>
    <category term="Technology Integration"/>
    <category term="Taxonomy"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/66de94e98e4755bbf2139eaf75d385646148801f-3018x1538.png?rect=136,0,2734,1538&amp;amp;w=1200&amp;amp;h=675&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;A single taxonomy structure shown on two screens, the first in a spreadsheet, and the second in a custom taxonomy management tool in Sanity Studio.. &quot; /&gt;&lt;/figure&gt;&lt;p&gt;Developing a taxonomy to meet the business needs of your content collection is a crucial step toward regaining control of your content ecosystem. Too often, however, well-intentioned taxonomy solutions never fully make it to the systems they&amp;#x27;re meant to serve. Don&amp;#x27;t let this be your fate! Here&amp;#x27;s the process I use to migrate taxonomy drafts in a spreadsheet to the taxonomy management tool in Sanity CMS.&lt;/p&gt;&lt;p&gt;In recent articles, I&amp;#x27;ve discussed &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/ai-taxonomy-term-generation/&quot;&gt;using AI to generate taxonomy terms&lt;/a&gt; and &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/purpose-driven-taxonomy-design/&quot;&gt;using a purpose-driven design process for shaping those terms into a taxonomy that meets business and user needs&lt;/a&gt;. None of this does any good, however, if your shiny new taxonomy sits moldering in a spreadsheet and never makes it to your production environment. (You would be alarmed by how often this happens. Alarmed.)&lt;/p&gt;&lt;p&gt;In this article I&amp;#x27;ll share the process I use to migrate a taxonomy from a spreadsheet to the taxonomy management tool in &lt;a href=&quot;https://www.sanity.io/&quot;&gt;Sanity CMS&lt;/a&gt;. I realize, of course, that not everyone&amp;#x27;s into Sanity. So it goes. While this example will focus on a Sanity implementation, the principles and core elements of the migration approach are not platform dependent. If you need a starting point for migrating taxonomy terms on a different platform, read on—I&amp;#x27;ll give guidance along the way on how you can adapt this process to other contexts.&lt;/p&gt;&lt;h2 id=&quot;embracing-standards&quot;&gt;Embracing Standards&lt;/h2&gt;&lt;p&gt;Whatever tech stack you use, embracing open, shared standards is a cornerstone of creating interoperable, reusable content ecosystems. The interoperability that comes with the use of standards is particularly important to the implementation of a taxonomy for (at least) two reasons.&lt;/p&gt;&lt;p&gt;First, taxonomies are living documents. To remain effective, they must change and evolve—sometimes growing, sometimes shrinking or consolidating to more effectively meet business needs. Embracing a standards-based approach helps you support this inherent change by making updates and evolution predictable and by allowing you to support your taxonomy with different tooling as your feature set, team, and business require.&lt;/p&gt;&lt;p&gt;Second, taxonomies often serve as the &amp;quot;connective tissue&amp;quot; between knowledge systems. Taxonomy is literally the shared vocabulary that allows different systems to &lt;em&gt;mean&lt;/em&gt; the same thing when they &lt;em&gt;say&lt;/em&gt; a particular thing. When your own taxonomy is based on predictable, shared standards, you&amp;#x27;ll be more likely to communicate directly with other standards-based applications, services, and vocabularies.&lt;/p&gt;&lt;h2 id=&quot;the-migration-script&quot;&gt;The Migration Script&lt;/h2&gt;&lt;p&gt;The &lt;a href=&quot;https://www.sanity.io/schemas/import-taxonomy-terms-b46af03f&quot;&gt;script&lt;/a&gt; we&amp;#x27;ll use to migrate terms from spreadsheet to CMS-based management tool relies on adherence to the W3C&amp;#x27;s &lt;a href=&quot;https://www.w3.org/TR/skos-reference/&quot;&gt;Simple Knowledge Organization System (SKOS)&lt;/a&gt; standard in both the &lt;a href=&quot;https://docs.google.com/spreadsheets/d/1eWaO_8pOdC7QwN4umnRnniEVFECMUc1b42FA3Ea7D0c/edit?gid=1890562484#gid=1890562484&quot;&gt;source spreadsheet&lt;/a&gt; and the &lt;a href=&quot;https://www.sanity.io/plugins/taxonomy-manager&quot;&gt;destination tool&lt;/a&gt;. This is just one example of how following standards can make moving between environments smooth and painless (or at least less painful). If I eventually need to move this taxonomy &lt;em&gt;out&lt;/em&gt; of Sanity in order to take advantage of the features of a standalone taxonomy management tool, working with a standards-based structure will make that move much less painful in turn.&lt;/p&gt;&lt;p&gt;At a high level, this script:&lt;/p&gt;&lt;ol&gt;&lt;li&gt;defines a concept scheme based on script settings&lt;/li&gt;&lt;li&gt;creates concepts by extracting preferred labels from the taxonomy hierarchy, mapping metadata to SKOS keys used by the &lt;a href=&quot;https://www.sanity.io/plugins/taxonomy-manager&quot;&gt;Taxonomy Manager Plugin&lt;/a&gt;, and adding `broader` concept relationships based on hierarchy&lt;/li&gt;&lt;li&gt;adds concepts to concept scheme array&lt;/li&gt;&lt;li&gt;creates a single transaction to import into the designated Sanity content store&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;The core logical problem that makes this script more than just &amp;quot;content migration&amp;quot; happens in step two. This is where we map the hierarchy from the end-user-friendly spatial layout of the spreadsheet to a set of relationships that represent the formal, standards-based relationships the Taxonomy Manager plugin expects (and enforces).&lt;/p&gt;&lt;pre&gt;&lt;code class=&quot;language-javascript&quot;&gt;// 3c. Map concepts to the hierarchy
  if (conceptLevel == 0) {
    // Top Concept
    hierarchy = [currentConceptRef] // reset hierarchy with new top concept
    previous = currentConceptRef // set previous to the current concept _ref for the next iteration
    topConceptSet = true
  } else if (conceptLevel == 1 &amp;amp;&amp;amp; topConceptSet == false) {
    // L1 concept with no top concept
    hierarchy = [currentConceptRef] // reset hierarchy with new concept
    previous = currentConceptRef
  } else if (
    conceptLevel - 1 == previous.level || // Child concept
    conceptLevel == previous.level || // Sibling concept
    conceptLevel &amp;lt; previous.level // Ancestor concept
  ) {
    mapChildConcept(
      hierarchy,
      currentConcept,
      broaderDetected,
      conceptLevel,
      currentConceptRef,
      toRemove,
    )
  } else if (conceptLevel &amp;gt; previous.level + 1) {
    // Orphan concept: two or levels lower than previous term
    throw new Error(&apos;Inconsistent Hierarchy: Orphan term.&apos;)
  }&lt;/code&gt;&lt;/pre&gt;&lt;p&gt;As this code block in the mapping function moves through each term in the supplied &lt;code&gt;.csv&lt;/code&gt;, it tracks which terms immediately precede it and assigns a level of hierarchy based on the relative position of each term. This allows us to detect child, sibling, and ancestor concepts, as well as new instances of Top Concepts or first-level concepts where no Top Concept is present. These relationships are then added as references to the &lt;code&gt;broader&lt;/code&gt; array of the term in question.&lt;/p&gt;&lt;p&gt;If you&amp;#x27;re importing your taxonomy into a different platform, the details of how you connect and upload to that system will change, but the basic structure of extracting the hierarchy and mapping it to a SKOS concept scheme should be adaptable—if not something you can just drop into your implementation (provided you&amp;#x27;re using standards-compliant tools, of course). Check out the &lt;a href=&quot;https://www.sanity.io/schemas/import-taxonomy-terms-b46af03f&quot;&gt;full script on Sanity Exchange&lt;/a&gt; for annotations and explanations of the additional functions and processes that support this core structure.&lt;/p&gt;&lt;h2 id=&quot;the-workflow&quot;&gt;The Workflow&lt;/h2&gt;&lt;p&gt;I explored &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/purpose-driven-taxonomy-design/&quot;&gt;purpose-driven taxonomy design&lt;/a&gt; in a recent article. That post provides a high level process for deciding what terms to include in your taxonomy and how to structure them. Once you&amp;#x27;ve defined, tested, and vetted your taxonomy, you can use these steps to migrate it into your instance of Sanity CMS.&lt;/p&gt;&lt;h3 id=&quot;preparing-the-taxonomy&quot;&gt;Preparing the Taxonomy&lt;/h3&gt;&lt;p&gt;The migration script is designed to work with a &lt;code&gt;.csv&lt;/code&gt; file exported from &lt;a href=&quot;https://docs.google.com/spreadsheets/d/1eWaO_8pOdC7QwN4umnRnniEVFECMUc1b42FA3Ea7D0c/edit?gid=1890562484#gid=1890562484&quot;&gt;this Google Sheets template&lt;/a&gt;. Download the &amp;quot;Taxonomy Template&amp;quot; tab (or whatever you choose to rename it) as a &lt;code&gt;.csv&lt;/code&gt; and add its location as the &lt;code&gt;sourceFile&lt;/code&gt;.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;Taxonomy Preparation Tips&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Use only one term per row. This is modeled in the spreadsheet sample data. The script will throw an error if there is more than one term per row or if there is no term specified for a row that otherwise contains metadata&lt;/li&gt;&lt;li&gt;Don&amp;#x27;t repeat terms. All your concept labels must be unique.&lt;/li&gt;&lt;li&gt;Use Top Concepts or L1 only, but not both. L1s will be nested under the nearest top concept if top concepts are used. Read more about &lt;a href=&quot;https://sanitytaxonomymanager.com/#/?id=adding-concepts&quot;&gt;Top Concepts in the Sanity Taxonomy Manager docs&lt;/a&gt;.&lt;/li&gt;&lt;li&gt;Include definitions! Definitions are helpful for your authors and indexers, as well as to any other &amp;quot;agents&amp;quot; (a-hem: 🤖) that may use your terms. Provide examples and scope notes, too, where appropriate. Read more about &lt;a href=&quot;https://www.w3.org/TR/2009/NOTE-skos-primer-20090818/#secdocumentation&quot;&gt;scope notes and examples in the W3C&amp;#x27;s SKOS Primer&lt;/a&gt;.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;script-settings&quot;&gt;Script Settings&lt;/h3&gt;&lt;p&gt;Once you have the &lt;code&gt;.csv&lt;/code&gt; saved and the path updated, add a title for the taxonomy scheme and an optional (but recommended) description.&lt;/p&gt;&lt;pre&gt;&lt;code class=&quot;language-javascript&quot;&gt;// 🚨 Change these to your values

const sourceFile = &apos;topic-taxonomy.csv&apos; // required; expects format provided by template linked in description
const conceptSchemeName = &apos;Topic Taxonomy&apos; // required
const conceptSchemeDescription = &apos;A taxonomy of topics used in Sanity Studio&apos; // optional
const baseIri = &apos;https://studio.sanity.io/taxonomy/&apos; // required; your studio domain is usually a good choice&lt;/code&gt;&lt;/pre&gt;&lt;p&gt;You&amp;#x27;ll also add a &lt;code&gt;baseIri&lt;/code&gt;. This is the default URI (Uniform Resource Identifier) for your concepts and concept schemes. Unique identifiers allow for the clear and unambiguous identification of concepts across namespaces, for example between &lt;code&gt;https://shipparts.com/vocab/Bow&lt;/code&gt; and &lt;code&gt;https://wrappingsupplies.com/vocab/Bow&lt;/code&gt;. The base URI of these concepts is &lt;code&gt;https://shipparts.com/vocab/&lt;/code&gt; and &lt;code&gt;https://wrappingsupplies.com/vocab/&lt;/code&gt;, respectively.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;In most cases, it makes sense for your base URI to be a directory or subdirectory of your website.&lt;/li&gt;&lt;li&gt;In all cases, the URI you choose should be in a domain that you control.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;The &lt;code&gt;baseIri&lt;/code&gt; can be changed later if you need to adjust namespaces after import.&lt;/p&gt;&lt;h3 id=&quot;running-the-script&quot;&gt;Running the Script&lt;/h3&gt;&lt;p&gt;To connect the script with your Sanity Content Lake instance, you&amp;#x27;ll need to include an &lt;code&gt;.env&lt;/code&gt; file with:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;your Sanity project ID&lt;/li&gt;&lt;li&gt;the name of your dataset&lt;/li&gt;&lt;li&gt;an API token with write access&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;With that in place and dependencies installed, run &lt;code&gt;node taxo-tool.js&lt;/code&gt; to migrate your taxonomy. Add flags to test or remove the entire scheme (and all its concepts):&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;--test&lt;/code&gt; prints the schema name, the first three concepts, and the total concept count to the terminal. No data is sent to Sanity.&lt;br/&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;--remove&lt;/code&gt; removes the taxonomy as configured in the script from your Studio. Use this if you need to make major changes and would rather start fresh.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;The script is idempotent, so if you update your schema and run it again, it will only change concepts that have been modified. Note that concept and scheme IDs are based on a hash of the concept preferred label (or the scheme title), so if you change labels or the scheme title, you will end up with duplicate entries. In this case you may just want to delete the entire scheme and migrate a fresh version.&lt;/p&gt;&lt;h2 id=&quot;next-steps&quot;&gt;Next Steps&lt;/h2&gt;&lt;p&gt;My hope is that this post helps those of you using Sanity—or another standards-compliant CMS—avoid the embarrassing fate of being stuck with taxonomy shelfware.&lt;/p&gt;&lt;p&gt;The next step, of course, is to start tagging content. In a future post, I&amp;#x27;ll explore some ways you can (responsibly) use LLMs to help with that. I&amp;#x27;ll also discuss how your structure and term definitions (you &lt;em&gt;are&lt;/em&gt; defining your terms, right? right??? 🧐) can help you create consistent, reliable results using purposefully coordinated, resource-appropriate models.&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">Purpose-Driven Taxonomy Design</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/purpose-driven-taxonomy-design/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/purpose-driven-taxonomy-design/</id>
    <published>2025-01-30T00:00:00Z</published>
    <updated>2025-01-30T00:00:00Z</updated>
    <summary type="text">A purpose-driven process for selecting terms, creating structure, and vetting design approaches on small and large scale taxonomy projects.</summary>
    <category term="Method"/>
    <category term="Taxonomy"/>
    <category term="Usability Testing"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/a6f9a65f1c739e90c526548398e13816d6f612b3-4566x1742.png?rect=1469,0,3097,1742&amp;amp;w=1200&amp;amp;h=675&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;A grid of digital sticky notes enumerating UX research and design themes&quot; /&gt;&lt;/figure&gt;&lt;p&gt;A purpose-driven taxonomy is designed to meet a clearly defined set of goals and use cases. This sounds obvious, but it&amp;#x27;s a step many organizations gloss over—often with frustratingly mixed outcomes as a result. In this article, I&amp;#x27;ll discuss a process I use for selecting terms, creating structure, and vetting design approaches on both small and large scale taxonomy projects. I&amp;#x27;ll also offer advice on where LLMs do—and don&amp;#x27;t—fit into a purpose-driven taxonomy design workflow.&lt;/p&gt;&lt;p&gt;In my last post, &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/ai-taxonomy-term-generation/&quot;&gt;AI Supported Taxonomy Term Generation&lt;/a&gt;, I used an LLM to help generate candidate terms for the revision of a topic taxonomy that had fallen out of sync with the content it was meant to tag. In that example, the taxonomy in question is for the &amp;quot;Insights&amp;quot; articles on &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/&quot;&gt;my consulting website&lt;/a&gt;. Having generated a set of term updates, the next step is now to decide which terms to include and then create an appropriate structure in which to organize them.&lt;/p&gt;&lt;p&gt;You might be thinking, &amp;quot;Great, throw it back into the LLM!&amp;quot; As many folks (&lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/taxonomy-boot-camp-2024-themes-takeaways/&quot;&gt;among them professional taxonomists&lt;/a&gt;) are figuring out, however, with this approach the devil is in the details. An LLM &lt;em&gt;may&lt;/em&gt; produce a syntactically correct taxonomy that &lt;em&gt;may&lt;/em&gt; conform to reality for a general topic, but you&amp;#x27;ll have to vet and validate each term to know if that&amp;#x27;s what it has done. Not only does correcting dodgy output not save you much time, but it also deprives you of the insight you gain by working through goals, options, and implementation mockups yourself.&lt;/p&gt;&lt;p&gt;This article is an overview of what that process looks like. I&amp;#x27;ll use my Topic Taxonomy revision as a running concrete example and will discuss along the way how these steps scale to larger projects with higher stakes—and more stakeholders. After an introduction to what I mean by &amp;quot;Purpose-Driven Taxonomy Design,&amp;quot; I&amp;#x27;ll discuss:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Defining taxonomy goals and use cases&lt;/li&gt;&lt;li&gt;Identifying candidate taxonomy terms&lt;/li&gt;&lt;li&gt;Defining relationships (creating the taxonomy&amp;#x27;s structure)&lt;/li&gt;&lt;li&gt;Moving the work into a spreadsheet to revise and refine&lt;/li&gt;&lt;li&gt;Vetting the proposed structure and terms&lt;/li&gt;&lt;li&gt;Migrating the taxonomy to a CMS&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;purpose-driven-design&quot;&gt;Purpose-Driven Design&lt;/h2&gt;&lt;p&gt;One common misstep I often see project teams make when beginning a taxonomy initiative is to build a taxonomy &amp;quot;of&amp;quot; a topic rather than &amp;quot;for&amp;quot; a purpose. A taxonomy &amp;quot;of&amp;quot; medical conditions, for instance, instead of a taxonomy &amp;quot;for&amp;quot; locating signs, symptoms, and when to see a physician. Or a taxonomy &amp;quot;of&amp;quot; courses offered, instead of &amp;quot;for&amp;quot; locating programs that align with a student&amp;#x27;s interests.&lt;/p&gt;&lt;p&gt;You can avoid this trap by first clearly identifying the goals and use cases your taxonomy needs to support, and then ensuring that the structure proposed meets those goals within the technology constraints at hand. This is a process I often follow when creating taxonomies for clients to support mission-critical operations.&lt;/p&gt;&lt;p&gt;While the tools, inputs, and scale needed to develop an enterprise taxonomy will differ from developing the small, low-stakes taxonomy I discuss here, the basic steps can be applied in both instances—and in both cases will usually result in a taxonomy that is more effective and easier to use, maintain, and iterate as your knowledge management needs evolve over time.&lt;/p&gt;&lt;h2 id=&quot;goals-use-cases&quot;&gt;Goals &amp;amp; Use Cases&lt;/h2&gt;&lt;p&gt;To design a taxonomy &lt;em&gt;for&lt;/em&gt; some intended outcome, you need to know what that outcome is. Yes, I know: this feels too obnoxiously obvious to put down in writing. But here we are. I&amp;#x27;ve seen enough cases of organizations skipping this step and ending up frustrated—or worse, feeling that taxonomy &amp;quot;doesn&amp;#x27;t work&amp;quot;—that I&amp;#x27;ll risk the annoyance.&lt;/p&gt;&lt;p&gt;In the case of my consulting website topic taxonomy, the &lt;em&gt;goals&lt;/em&gt; of the taxonomy are to:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;coordinate content recommendations&lt;/li&gt;&lt;li&gt;communicate the range of topics I write about&lt;/li&gt;&lt;li&gt;afford topic-based browsing&lt;/li&gt;&lt;li&gt;complement site traffic data to show which topics and categories generate the most interest among site visitors&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;These are the outcomes I want the taxonomy to help me achieve. It&amp;#x27;s useful to articulate these, and, if there are multiple stakeholders, align on them. Different goals will suggest different structures and terms. I &lt;em&gt;don&amp;#x27;t&lt;/em&gt; need this taxonomy to support faceted browsing, for instance. It&amp;#x27;s also not a priority (right now) for it to include alternate and non-preferred terms to support search indexing and retrieval.&lt;/p&gt;&lt;p&gt;The &lt;em&gt;use cases&lt;/em&gt; of this taxonomy describe how it will be employed to achieve these goals. For this example, my topic taxonomy needs to:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;support the assignment of multiple tags to accurately represent article topics&lt;/li&gt;&lt;li&gt;support the association of articles that share an identical tag&lt;/li&gt;&lt;li&gt;support the association of articles that share &lt;em&gt;sibling&lt;/em&gt; tags&lt;/li&gt;&lt;li&gt;support the association of articles that share &lt;em&gt;cousin&lt;/em&gt; tags&lt;/li&gt;&lt;li&gt;support end-user browsing by assigned tags&lt;/li&gt;&lt;li&gt;support end-user browsing by all tags&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;For these use cases, associations will be ranked based on how close the relationship between terms is. This means that the specificity of terms should be consistent at different levels of the eventual hierarchy so that terms at similar levels return a predictable number and specificity of results.&lt;/p&gt;&lt;h2 id=&quot;candidate-terms&quot;&gt;Candidate Terms&lt;/h2&gt;&lt;p&gt;With a clear sense of why this taxonomy exists and what it needs to accomplish, I can now make informed decisions about which terms to include. In my &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/ai-taxonomy-term-generation/&quot;&gt;last article&lt;/a&gt; I showed how to use an LLM to generate a ranked list of candidate terms for a collection of content. I&amp;#x27;ll narrow this list down by choosing terms that are most likely to help me meet the goals defined above for the intended audience of the content and the purpose of the site as a whole.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/b54d0411e7af115190eedd5eaa5de35a27598f26-3023x1725.webp?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;685&quot; alt=&quot;A spreadsheet with a list of recommended tags in the left column, conditionally highlighted statistics in the next three columns, and synthesized definitions for each&quot; /&gt;&lt;figcaption&gt;LLM-generated candidate topic tags.&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Each term returned from the term generation process includes data for:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;the number of resources to which it was applied&lt;/li&gt;&lt;li&gt;the average relevance score of the tag to its content&lt;/li&gt;&lt;li&gt;the standard deviation of the relevance score&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;This data allows me to first rank terms by those with the highest average relevance and easily identify those with the lowest standard deviation of relevance scores. This means that for most of the content tagged, the relevance was similarly high.&lt;/p&gt;&lt;p&gt;&amp;quot;Information Architecture,&amp;quot; &amp;quot;Content Strategy,&amp;quot; and &amp;quot;Semantic Web,&amp;quot; for example, all have high relevance scores and low standard deviation. The first two of those terms, however, can be applied to most (if not all) articles in the collection. This doesn&amp;#x27;t do much to help with the goal of affording topic-based browsing. A category of &amp;quot;all content&amp;quot; doesn&amp;#x27;t narrow the set down. But they could be good candidate terms for first or second level categories.&lt;/p&gt;&lt;p&gt;Tags with high relevance and a high standard deviation need investigation. By looking at the Raw Tags data returned from the LLM, I can investigate why the range of relevancy scores is broad. For &amp;quot;taxonomy,&amp;quot; for instance, two articles have a low score because taxonomy is part of what they discuss, but isn&amp;#x27;t the main topic of the article. This supports the content recommendation and topic-based browsing goals discussed above.&lt;/p&gt;&lt;p&gt;Some tags, however, like &amp;quot;digital marketing,&amp;quot; have a low standard deviation (0.035) and a reasonable measure of relevance (0.75) but are only used twice. These are relatively weak connections between a small set of resources. They also don&amp;#x27;t align strongly with the site&amp;#x27;s purpose. I&amp;#x27;ll likely leave terms like these out of the first draft.&lt;/p&gt;&lt;h2 id=&quot;relationships&quot;&gt;Relationships&lt;/h2&gt;&lt;p&gt;Once I have a set of candidate terms, it&amp;#x27;s time to sort out how they relate to one another. Like the previous step, this one is driven by the goals and use cases defined for the taxonomy, the needs of the site audience, and the purpose of the content collection.&lt;/p&gt;&lt;p&gt;For larger projects with multiple stakeholders and audiences, card sorting with your content set&amp;#x27;s target audience is a great way to gather evidence for key relationships and structures to consider. Few techniques are as simple to set up and rich in the insights they deliver about users&amp;#x27; mental models.&lt;/p&gt;&lt;p&gt;For this small, low-stakes example, sketching structural approaches in a white-boarding tool like Miro is a good way to explore possible options. I started this process by deduplicating concepts and color-coding them into simple categories based on the candidate term selection step. This helped me keep track of where terms originated and how &amp;quot;important&amp;quot; they are to the overall content set.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/117ead8c69bd479490095f01c5903dbe094a96d2-5120x2880.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;A Miro board showing a grid of sticky notes in three different colors, each with a different taxonomy term written on them&quot; /&gt;&lt;figcaption&gt;Deduplicated candidate terms&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;I then grouped together similar terms, assigning parent terms to groups of similar concepts and explored ways to fit the whole structure together into a single hierarchy. &lt;/p&gt;&lt;p&gt;The approach that initially seemed the most obvious to me was to organize by discipline (Information Architecture, Content Strategy, User Experience Design, etc.), but trying this out quickly showed a problem: many of the tags I had selected belonged in several of these categories at once.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/a34b5de59426ff31eef4f3b9e076dec0abdee272-3366x2242.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;799&quot; alt=&quot;A collection of sticky notes organized in an incomplete hierarchical structure in an online white-boarding tool. &quot; /&gt;&lt;figcaption&gt;Candidate terms organized by discipline&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;While this arrangement is possible—it&amp;#x27;s an approach called &amp;quot;polyhierarchy&amp;quot;—it creates a complex set of relationships to manage. Without a compelling reason for the complexity, it&amp;#x27;s best avoided—particularly as a starting point.&lt;/p&gt;&lt;p&gt;The next approach I tried was to organize terms by &amp;quot;project phase focus.&amp;quot; This strategy allowed me to eliminate the polyhierarchy and resulted in a more balanced set of second and third level categories from the start.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/513fe5482522a5ed62acfe1eaf38338cf6ea39eb-4576x2184.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;573&quot; alt=&quot;A collection of sticky notes organized in a hierarchical structure in an online white-boarding tool. &quot; /&gt;&lt;figcaption&gt;Candidate terms organized by project phase focus&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;This approach also highlighted &lt;em&gt;gaps&lt;/em&gt; in my candidate term set. As I sorted and categorized, I filled in missing concepts from my knowledge of the domain, my content set, my business goals, and my users. These were tags that either provided organizational affordances or weren&amp;#x27;t currently represented in my content set—but which I could reasonably anticipate would be useful for future articles and case studies.&lt;/p&gt;&lt;p&gt;Though this second structure felt like a good starting point, I explored one more approach to stretch my thinking. This third option uses a &amp;quot;discipline focus by phase&amp;quot; strategy that swaps the first and second level categories.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/a3d3bde4d12b4d44bf88b5d5384cff39f12dce6e-4646x2334.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;603&quot; alt=&quot;A collection of sticky notes organized in a hierarchical structure in an online white-boarding tool. &quot; /&gt;&lt;figcaption&gt;Candidate terms organized by discipline focus phase&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;This version isn&amp;#x27;t the worst thing in the world, but it creates duplication at level two that would require longer (more specified) labels for those terms. It is also less balanced than the &amp;quot;project phase by focus&amp;quot; approach. This would mean that tags in some branches could artificially create more connections than others, just by virtue of their structure.&lt;/p&gt;&lt;p&gt;As neither of these features works to support my goals and use cases, &amp;quot;project phase by focus&amp;quot; is the model I&amp;#x27;ll move forward with. If you&amp;#x27;re collaborating with other designers, engineers, or business stakeholders, vetting your selection at this step in a visual tool can be a great way to gather feedback. Virtual whiteboards and sticky notes can be easier for some audiences to handle than spreadsheets—which is where we&amp;#x27;re heading next.&lt;/p&gt;&lt;h2 id=&quot;define-revise-and-refine&quot;&gt;Define, Revise, and Refine&lt;/h2&gt;&lt;p&gt;While digital white-boarding tools like Miro can be a great place to draft structural approaches, the most common tool for revising and vetting a taxonomy is the ubiquitous spreadsheet. Spreadsheets allow you to easily share and iterate your draft, and, if properly formatted, give you a structure you can import into a standards-compliant taxonomy management tool.&lt;/p&gt;&lt;p&gt;A note on &amp;quot;proper formatting&amp;quot;: make sure each record (in this case, each taxonomy term) is in its own row. This is a convention I often see skipped, and it makes migrating your structure to a production tool much more complicated than it needs to be. For a primer on data structures for designers, see Tim Sheiner&amp;#x27;s excellent article &amp;quot;&lt;a href=&quot;https://medium.com/salesforce-ux/designers-field-guide-to-data-a573693cdea8&quot;&gt;Understanding Data&lt;/a&gt;.&amp;quot; To make your own blank copy of the spreadsheet I use below, check out the &lt;a href=&quot;https://docs.google.com/spreadsheets/d/1eWaO_8pOdC7QwN4umnRnniEVFECMUc1b42FA3Ea7D0c/edit?usp=sharing&quot;&gt;Taxonomy Manager Google Sheets template here&lt;/a&gt;.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/9c314f25d5875dbed68100e0a8215be5b01fd7b6-3026x1722.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;683&quot; alt=&quot;A spreadsheet showing a hierarchical organization of terms and term definitions associated with each.&quot; /&gt;&lt;figcaption&gt;Candidate taxonomy terms, relationships, and definitions in the template worksheet&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;To get my candidate terms into the template, I highlight the relevant set of &amp;quot;sticky notes&amp;quot; in Miro and paste them into the first column. I then organize them into the intended hierarchy and use a &lt;code&gt;VLOOKUP()&lt;/code&gt; function in the &amp;quot;Definitions&amp;quot; column to bring over generated definitions from the &amp;quot;Processed Tags&amp;quot; sheet:&lt;/p&gt;&lt;pre&gt;&lt;code&gt;=VLOOKUP(C5,&apos;Processed Tags&apos;!A$2:G$39,7,FALSE)&lt;/code&gt;&lt;/pre&gt;&lt;p&gt;🚨 Do not take these definitions at face value! 🚨 &lt;/p&gt;&lt;p&gt;Review, vet, and refine the LLM-generated definitions, and create definitions for any terms you added in the sketching phase. While it&amp;#x27;s possible to outsource this step to an LLM, the mental activity of creating definitions for this context is a powerful tool for vetting and refining your structure. I find that defining the terms also helps me spot inconsistencies and iterate on the draft approach to create a stronger, more cohesive whole.&lt;/p&gt;&lt;p&gt;This is also a step where, if I&amp;#x27;m working with project stakeholders and collaborators, I&amp;#x27;ll meet with them to get their feedback and input. Even if you&amp;#x27;re working alone, however, careful attention to the definition phase can help you avoid frustrating errors.&lt;/p&gt;&lt;h2 id=&quot;vet-the-proposed-structure&quot;&gt;Vet the Proposed Structure&lt;/h2&gt;&lt;p&gt;Once you have a revised structure that is coherent, internally consistent, and representative of your goals, test it out against your use cases. As with previous steps, if you&amp;#x27;re working on a large taxonomy with multiple stakeholders, there are time-tested approaches for getting feedback at scale. Closed card sorting and tree testing are both excellent methods for vetting that a proposed structure is consistent with intended users&amp;#x27; mental models.&lt;/p&gt;&lt;p&gt;At a small scale, manual approaches to vetting a structure against use cases can provide an effective sanity check. For the topic taxonomy discussed here, I pulled five recent articles from my website, tagged them with terms from the proposed taxonomy draft, and then traced the connections between the assigned tags based on the association rules outlined in my use cases.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/5eedb962e760ebdf3a0ede5455244781c0efc302-3340x2808.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;1009&quot; alt=&quot;A virtual whiteboard display of six article previews &amp;quot;tagged&amp;quot; with term sticky notes, and &amp;quot;related&amp;quot; with colored arrows. Expected &amp;quot;related articles&amp;quot; recommendations are listed next to each. &quot; /&gt;&lt;figcaption&gt;Test articles manually tagged with candidate terms&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;For this exercise, I drew red connectors for tag-to-tag matches and orange connectors for sibling matches. Solid lines represent connections between the first three ranked tags; dashed lines are connections to a lower-priority (less relevant) tag. The result is messy, but it&amp;#x27;s fast and helps you understand what results your proposed rules and structure will create.&lt;/p&gt;&lt;p&gt;Showing these connections and weights visually allows me to approximate which related articles the tags assigned to a given piece of content are likely to produce. If these are the relationships I expect—or what I would suggest if I were manually connecting content—I&amp;#x27;m likely on the right track. If not, I&amp;#x27;ll return to the model and look for sources of the mismatch.&lt;/p&gt;&lt;p&gt;This is another instance where if you&amp;#x27;re working with stakeholders or subject matter experts, you&amp;#x27;ll want to involve them in reviewing, validating, and nuance these findings.&lt;/p&gt;&lt;h2 id=&quot;migrate-to-the-cms&quot;&gt;Migrate to the CMS&lt;/h2&gt;&lt;p&gt;Now that I&amp;#x27;ve vetted my terms and structure against my use cases and mocked up a sample tagging scenario, I can be reasonably confident that the taxonomy I&amp;#x27;ve created will support the goals I defined at the outset. The true test of a plan, however, is contact with the real world: It&amp;#x27;s time to migrate the taxonomy to my content management system.&lt;/p&gt;&lt;p&gt;For any but the smallest taxonomies, manually migrating terms can be a tedious and error-prone process. Fortunately, if you&amp;#x27;ve followed a standardized process in creating your structure, migrating terms and relationships to a standards-compliant taxonomy management tool needn&amp;#x27;t be an ordeal&lt;/p&gt;&lt;p&gt;The spreadsheet template I used to create the topic taxonomy above adheres to the &lt;a href=&quot;https://www.w3.org/TR/skos-reference/&quot;&gt;Simple Knowledge Organization System (SKOS)&lt;/a&gt; standard established by the World Wide Web Commission (W3C). The taxonomy management tool to which I&amp;#x27;ll be migrating, the &lt;a href=&quot;https://www.sanity.io/plugins/taxonomy-manager&quot;&gt;Sanity Studio Taxonomy Manager&lt;/a&gt; plug-in (which, coincidentally, I built and maintain), is also SKOS compliant. Thanks to Sanity.io&amp;#x27;s integration-friendly content store, I can map and upload the entire taxonomy with a single node.js script.&lt;/p&gt;&lt;p&gt;For those of you already using Sanity, I&amp;#x27;ll share more details on using this script in a follow-up post. If you&amp;#x27;re using a different CMS—or are evaluating tech stack options—it&amp;#x27;s well worth researching the degree to which your tools support open standards and free movement between platforms. Just as your taxonomy should be structured (and sized) to support goals that meet clearly defined outcomes, the tools you use to deploy and govern that taxonomy should likewise be chosen to support its scale and interoperability requirements.&lt;/p&gt;&lt;p&gt;Once your taxonomy is integrated with your CMS, the next step, of course, is to start tagging content. In a future post I&amp;#x27;ll explore some ways you can (responsibly) use LLMs to help with that. I&amp;#x27;ll also discuss how your structure and term definitions (you &lt;em&gt;are&lt;/em&gt; defining your terms, right? 🧐) can help you create consistent, reliable results using purposefully coordinated, resource-appropriate models.&lt;/p&gt;</content>
  </entry>
  <entry>
    <title type="text">AI Supported Taxonomy Term Generation</title>
    <link rel="alternate" type="text/html" href="https://preview.andyfitzgeraldconsulting.com/insights/ai-taxonomy-term-generation/"/>
    <id>https://preview.andyfitzgeraldconsulting.com/insights/ai-taxonomy-term-generation/</id>
    <published>2025-01-15T00:00:00Z</published>
    <updated>2025-01-15T00:00:00Z</updated>
    <summary type="text">An example workflow for coordinating LLMs and text analytics to generate topic tags for an article collection.</summary>
    <category term="Method"/>
    <category term="Taxonomy"/>
    <category term="AI Integration"/>
    <category term="Python"/>
    <content type="html">&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/b20d2045c9b611e76678883e7cecf66196ec4ad8-3573x2311.png?rect=0,186,3573,2010&amp;amp;w=1200&amp;amp;h=675&amp;amp;fit=crop&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;675&quot; alt=&quot;Diamond patterned wire fence with small circular metal tags printed with names hanging from each vertice&quot; /&gt;&lt;/figure&gt;&lt;p&gt;Underneath the hype and hyperbole of &amp;quot;AI&amp;quot; capabilities, there&amp;#x27;s a compelling use case for responsibly using LLMs to help information architecture practitioners make better use of time-tested analysis and classification approaches and tools. Here is an example of one way that LLMs and rules-based text analysis techniques can be coordinated to generate tags for a topic taxonomy that has fallen out of sync with its content set. &lt;/p&gt;&lt;p&gt;Slowly but surely, &amp;quot;AI&amp;quot; in the form of large language models (LLMs) is becoming a part of my information architecture toolkit. As I&amp;#x27;ve &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/deskilling-large-language-models/&quot;&gt;written about elsewhere&lt;/a&gt;, I find they&amp;#x27;re most useful for the kind of IA work I do for my clients when scoped to narrow, specific tasks, often in a workflow with other natural language processing, logic-based reasoning, and human evaluation steps.&lt;/p&gt;&lt;p&gt;A recent experiment I conducted with my consulting website provides an example of what this can look like in practice. Like many websites that have been around for more than a few years (and that may or may not be guilty of a lack of diligence with their governance process 😬), I found that the terms in my topic taxonomy had fallen out of sync with the kinds of topics I now write about.&lt;/p&gt;&lt;p&gt;I &lt;em&gt;could&lt;/em&gt; re-read all of my articles, posts, and case studies and manually generate a revised list of candidate topic tags. I have just enough content, however, to make the prospect of going through it piece by piece daunting. I likewise don&amp;#x27;t sufficiently trust any of the current generation of language models (or &amp;quot;graph rag&amp;quot; systems) to capture the user, content, and business aspects of my taxonomy needs to hand over the task wholesale.&lt;/p&gt;&lt;p&gt;Properly guided and supervised, however, an LLM can be one element of a coordinated toolchain for this task. This article is a walkthrough of one way that chain can be assembled. What follows is a look at:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;information design considerations for toolchain steps&lt;/li&gt;&lt;li&gt;key toolchain tasks&lt;/li&gt;&lt;li&gt;results produced &amp;amp; next steps&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;👉 View the &lt;a href=&quot;https://github.com/andybywire/tag-gen&quot;&gt;full repo for this workflow on GitHub&lt;/a&gt;.&lt;/p&gt;&lt;h2 id=&quot;goal&quot;&gt;Goal&lt;/h2&gt;&lt;p&gt;The goal of this workflow is &lt;em&gt;not&lt;/em&gt; to tag content. Free-tagging content and tagging content with existing taxonomy terms are workflows that have already gotten a lot of attention, and there are alreacy many excellent examples of how to accomplish this out there.&lt;/p&gt;&lt;p&gt;I also don&amp;#x27;t want an LLM to generate a taxonomy &lt;em&gt;for&lt;/em&gt; me. A well-designed taxonomy affords a nuanced expression of user needs, business goals, and content focus in a specific context. This requires judgment and accountability. LLMs are notorious for providing answers that &lt;em&gt;look&lt;/em&gt; right, but may or may not be garbage. As taxonomists &lt;a href=&quot;https://conferences.infotoday.com/documents/458/1215_Moore.pdf&quot;&gt;Shannon Moore and Max Gaibort from Electronic Arts&lt;/a&gt; so aptly put it in their presentation at &lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/taxonomy-boot-camp-2024-themes-takeaways/&quot;&gt;Taxonomy Boot Camp&lt;/a&gt; in 2024, &amp;quot;taxonomy-shaped responses are not a taxonomy.&amp;quot;&lt;/p&gt;&lt;p&gt;I do, however, want a qualified set of candidate terms that &lt;em&gt;could&lt;/em&gt; work as entries in my revised taxonomy. &amp;quot;Qualified,&amp;quot; in this sense, means those terms that have relevance to my users, content, and business goals (&amp;quot;warrants&amp;quot; in taxonomy lingo), and those that are sufficiently represented across my content collection that they will work to coordinate resources. A tag that only appears once—or can be applied to &lt;em&gt;every&lt;/em&gt; resource—is not helpful for discovery, recommendation, or analytics.&lt;/p&gt;&lt;p&gt;A qualified candidate tag should also include a definition and a context-sensitive rationale for why it applies to a particular resource. It should also only occur once in the proposed candidate terms list. This will give the eventual taxonomist (i.e. me) a set of useful inputs with which to make informed decisions about if and how a candidate term fits into the final revised taxonomy.&lt;/p&gt;&lt;h2 id=&quot;workflow-steps&quot;&gt;Workflow Steps&lt;/h2&gt;&lt;p&gt;An LLM is clearly a good fit for some of the goals defined above ... and would clearly be rubbish (or wantonly wasteful) for others. I also don&amp;#x27;t need the same &amp;quot;strength&amp;quot; of LLM for each generative task. Here&amp;#x27;s the approach I took to coordinate constituent steps so that each is completed with the best tool for the job.&lt;/p&gt;&lt;p&gt;👉 Follow along in the &lt;a href=&quot;https://github.com/andybywire/tag-gen/blob/main/tag_generator.ipynb&quot;&gt;full Notebook here&lt;/a&gt;.&lt;/p&gt;&lt;h3 id=&quot;1-fetch-tidy-articles&quot;&gt;1. Fetch &amp;amp; Tidy Articles&lt;/h3&gt;&lt;p&gt;I start by querying my CMS (built with &lt;a href=&quot;https://www.sanity.io/&quot;&gt;Sanity&lt;/a&gt;) for the relevant set of articles that should inform my generated tags. Sanity&amp;#x27;s &amp;quot;projections&amp;quot; allow me to return just the data I need and format keys to values that will be easy to reference later on.&lt;/p&gt;&lt;p&gt;In this GROQ query, for example, &lt;code&gt;type&lt;/code&gt; (sans underscore) is a reference to a term in my site&amp;#x27;s &amp;quot;types&amp;quot; taxonomy. The GROQ dereferencing operator (&lt;code&gt;-&amp;gt;&lt;/code&gt;) allows me to return just the preferred label for that term.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/82b2b00ea1f281668748e3500061565851a4fae6-1236x628.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;610&quot; alt=&quot;A response variable assigned to the output of a post request generated with the POST method. The query returns the title, URL, and type of each document of type &amp;quot;Article&amp;quot; or &amp;quot;Case Study&amp;quot;&quot; /&gt;&lt;figcaption&gt;A simple GROQ query for &amp;quot;article&amp;quot; and &amp;quot;case study&amp;quot; content types&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;I use &lt;a href=&quot;https://beautiful-soup-4.readthedocs.io/en/latest/&quot;&gt;Beautiful Soup&lt;/a&gt; to strip out decorative headers, timestamps, and existing lists of tags and topics from the returned results. I then convert the HTML to Markdown to make it easier for the LLM to process. This gives the language model a representation similar to what site visitors see when visiting article pages.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/7b7ee65fbb560296cd8da7651b035a009be8951a-1164x1318.png?w=1164&amp;amp;auto=format&quot; width=&quot;1164&quot; height=&quot;1318&quot; alt=&quot;A Python `for` loop that removes topic tags, decorative headers, and timestamps from article HTML code. The resulting code is subsequently transformed into Markdown. &quot; /&gt;&lt;figcaption&gt;Removing existing tags with Beautiful Soup and converting to Markdown&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;All of this content is then added to a Pandas DataFrame. I also add in the word count of each final article, which I&amp;#x27;ll use later to determine a range for the number of tags the LLM should produce.&lt;/p&gt;&lt;h3 id=&quot;2-create-topic-tag-generation-prompts&quot;&gt;2. Create Topic Tag Generation Prompts&lt;/h3&gt;&lt;p&gt;Now that I&amp;#x27;ve got the content data I need an accessible format, it&amp;#x27;s time to give the LLM some guidance. I don&amp;#x27;t want it to &amp;quot;free tag&amp;quot; this content in the broad sense of the term. My site has a narrow purpose and is intended to speak to a specific audience. I want the LLM to take that into account.&lt;/p&gt;&lt;p&gt;I accomplish this with a system prompt structured into five sections (&lt;a href=&quot;https://github.com/andybywire/tag-gen/blob/main/tag_generator.ipynb&quot;&gt;see the full prompts in the Jupyter Notebook&lt;/a&gt;):&lt;/p&gt;&lt;ol&gt;&lt;li&gt;A general overview of what the requested task is and what it will accomplish&lt;/li&gt;&lt;li&gt;A statement about the purpose of the AF Consulting website&lt;/li&gt;&lt;li&gt;A concise set of target audience profiles&lt;/li&gt;&lt;li&gt;The article to be tagged&lt;/li&gt;&lt;li&gt;A statement describing the grammatical, orthographic, and metadata structure of the tags to be returned&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;Because this is a relatively complex task with a fairly large input, I&amp;#x27;ve opted for a hosted LLM to carry it out (GPT 4o-mini). For consistency, I also want a model that can enforce structured output. In this case, the model will return a Python dictionary with each output element stored as a value in a predefined structure that I can easily iterate over later.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/db622c5de85aa5d51b3ff11b505e0da16d6b7553-1228x1140.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;1114&quot; alt=&quot;A Pydantic class specifying that the generated tag output should be returned as a set of keys with names: tag, explanation, and relevance. &quot; /&gt;&lt;figcaption&gt;The structured output definition for the LLM tag generation task. &lt;/figcaption&gt;&lt;/figure&gt;&lt;h3 id=&quot;3-generate-tags&quot;&gt;3. Generate Tags&lt;/h3&gt;&lt;p&gt;To generate tags, I loop through the article content in the &amp;quot;articles&amp;quot; DataFrame and submit each with the above prompt to the language model. This step is also where I adjust the number of tags requested based on the length of the article:&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/5560b11cea7525cedebf49e358cf3f99560af4e8-1170x166.png?w=1170&amp;amp;auto=format&quot; width=&quot;1170&quot; height=&quot;166&quot; alt=&quot;An expression setting the count variable to 10 - 20 tags, depending on the length of the article&quot; /&gt;&lt;figcaption&gt;The LLM should return 10 - 20 tags, depending on the length of the article&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;To be clear, 10 - 20 tags are far too many tags for any single article. Just as I would request of a human tagger, however, I want the model to exhaust the &amp;quot;obvious&amp;quot; tags and dig for potential outliers. Most of these will be thrown out, but by collecting and reviewing them, I keep the process open to serendipity and the discovery of new ways to describe existing content—or new angles to explore in future articles.&lt;/p&gt;&lt;p&gt;After this expansive pass of tags is generated, I export it as a &lt;code&gt;.csv&lt;/code&gt;. This allows me to reuse a copy of this step&amp;#x27;s output if I need to iterate subsequent steps (and saves on the resource consumption associated with a hosted LLM). It also provides a reference output for the taxonomy authoring work that comes in the next step.&lt;/p&gt;&lt;h3 id=&quot;4-process-tag-list&quot;&gt;4. Process Tag List&lt;/h3&gt;&lt;p&gt;Though useful to generate, the broad take on topic tags generated in the previous step is not the primary input I want to work with when I take a first pass at deciding which tags should be included in the revised taxonomy. Because I&amp;#x27;ve also collected &amp;quot;relevance&amp;quot; scores along the way, I can use this value to filter out low relevance tags that only occur once:&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/d4aecbcb4f488746f36b8f0a2e220e400091da5e-1238x252.png?w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;244&quot; alt=&quot;A filter method with a lambda function that only includes tags with an article count of more than one and a relevance score greater than 0.6&quot; /&gt;&lt;figcaption&gt;Singleton and low-relevance tags are removed from the filtered set&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Low-relevance tags that occur in multiple places may have a story to tell, so I&amp;#x27;ll leave those be.&lt;/p&gt;&lt;p&gt;This step is also where I combine duplicate entries. As you&amp;#x27;d imagine, &amp;quot;information architecture&amp;quot; and &amp;quot;content strategy&amp;quot; occur more than once. When deduplicating tags I also calculate a set of stats from the set:&lt;/p&gt;&lt;ul&gt;&lt;li&gt;number of resources tagged&lt;/li&gt;&lt;li&gt;average &amp;quot;relevance&amp;quot; score&lt;/li&gt;&lt;li&gt;standard deviation of relevance scores&lt;/li&gt;&lt;li&gt;number of content types represented&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;These numbers help me understand at a glance how a particular tag fits with the content set as a whole. I also join together all of the definitions provided (I&amp;#x27;ll synthesize them in the next step).&lt;/p&gt;&lt;h3 id=&quot;5-synthesize-definitions&quot;&gt;5. Synthesize Definitions&lt;/h3&gt;&lt;p&gt;Tidying up the concatenated list of definitions created in the last step is another task well-suited to an LLM. Condensing and summarizing verbose content is an area where language models excel—and where humans quickly go crosseyed from the tedium of the work.&lt;/p&gt;&lt;p&gt;This step does not, however, require a high-power hosted LLM. Most of the concatenated definitions are not very long and they mostly repeat the same ideas. By using a smaller model hosted on my MacBook Pro, I can keep the cost and the resource impact of using an LLM to a minimum.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/34de57cc410157924d718c00c49b498647bf00cd-1234x1670.png?rect=0,0,1234,1560&amp;amp;w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;1517&quot; alt=&quot;An LLM prompt instructing the model to synthesize long, repetitive definitions into a single concise definition. &quot; /&gt;&lt;figcaption&gt;The definition synthesis prompt and loop&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;Synthesized definitions are added to a new column in the DataFrame. Unchanged definitions (i.e. for single-entry tags) are passed through to this column for consistency. I then export the trimmed, processed, and synthesized tags as a second &lt;code&gt;.csv&lt;/code&gt; for taxonomy development. If anything looks fishy in these final results, I can always refer to the full tag list exported in step three to compare individual entries.&lt;/p&gt;&lt;h3 id=&quot;6-visualize-results&quot;&gt;6. Visualize Results&lt;/h3&gt;&lt;p&gt;This step isn&amp;#x27;t strictly necessary, but it&amp;#x27;s a satisfying wrap-up for the workflow—and the visual representation offers a novel perspective on the generated tag collection as a whole.&lt;/p&gt;&lt;p&gt;For this dataset I&amp;#x27;ve configured a plot to show resource count and relevance score on the x-y axis and standard deviation as a color gradient. The standard deviation metric allows me to quickly identify tags that have the same—or very different—relevance to their respective articles. The size of the point shows the number of content types represented by each tag. Additional data point details are available on hover.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/8615fdd808131caf62b389d5e5b88c42ea441c55-1566x1540.png?rect=0,5,1566,1535&amp;amp;w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;1176&quot; alt=&quot;A plot with &amp;quot;number of resources&amp;quot; on the y axis and &amp;quot;relevance score&amp;quot; on the x axis. A color scale indicates the standard deviation of relevance values for each point.&quot; /&gt;&lt;figcaption&gt;An interactive plot helps visualize generated results&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;By exporting this final plot as an HTML file, I can preserve the interactive features for later reference. &lt;/p&gt;&lt;h2 id=&quot;results-next-steps&quot;&gt;Results &amp;amp; Next Steps&lt;/h2&gt;&lt;p&gt;If you&amp;#x27;ve grown accustomed to the &amp;quot;LLM as one-stop-shop&amp;quot; narratives that often play out on social media, you might be underwhelmed by the output of this process: two &lt;code&gt;.csv&lt;/code&gt; files and an interactive chart.&lt;/p&gt;&lt;p&gt;Please recall, however, that the goal of this workflow isn&amp;#x27;t to create a taxonomy—or attempt to reproduce the judgment calls made by a skilled taxonomist. The goal is to provide a taxonomist with &lt;em&gt;some&lt;/em&gt; of the input they need to create an effective vocabulary and help them (in this case, me) minimize the slog of manually reviewing material, exhausting tag options, and defining and rationalizing each candidate tag.&lt;/p&gt;&lt;p&gt;By importing the exported &lt;code&gt;.csv&lt;/code&gt; documents into a spreadsheet (taxonomy&amp;#x27;s natural spawning ground), I&amp;#x27;ve got an informed and annotated set of candidate terms in the environment in which taxonomies invariably take initial shape.&lt;/p&gt;&lt;figure&gt;&lt;img src=&quot;https://cdn.sanity.io/images/7v0qvet6/production-26/99372cfc2205ff66ce2c4fdab8af76584e3496c4-2844x1778.png?rect=0,0,2844,1774&amp;amp;w=1200&amp;amp;auto=format&quot; width=&quot;1200&quot; height=&quot;749&quot; alt=&quot;A spreadsheet with a list of recommended tags in the left column, conditionally highlighted statistics in the next three columns, and synthesized definitions for each&quot; /&gt;&lt;figcaption&gt;Light conditional formatting highlights patterns in tag results&lt;/figcaption&gt;&lt;/figure&gt;&lt;p&gt;I&amp;#x27;m able to quickly see, for instance, which tags are applicable to a majority of my content. I may consider these as top level terms—or may opt to exclude them altogether if they don&amp;#x27;t do any work to differentiate topics.&lt;/p&gt;&lt;p&gt;The standard deviation of relevance scores likewise helps me identify which tags may have different &amp;quot;strengths&amp;quot; in relation to a given article. In some cases, a parent term may be a better fit. This will inform the structure of the eventual hierarchy.&lt;/p&gt;&lt;p&gt;In the &amp;quot;Raw Tags&amp;quot; tab I can also see which articles had the most &amp;quot;rejected&amp;quot; tags. In this case, that was &amp;quot;&lt;a href=&quot;https://preview.andyfitzgeraldconsulting.com/insights/cognitive-science-for-designers/&quot;&gt;A Cognitive Sciences Reading List for Designers&lt;/a&gt;,&amp;quot; which makes sense as this post covers a dozen topics that inform my practice but which aren&amp;#x27;t a focus of it.&lt;/p&gt;&lt;p&gt;In a broad sense, these results provide data for an informed revision process that can take shape based on the needs of a particular project. If you&amp;#x27;re creating a completely new structure, for example, a card sorting exercise with candidate terms may be the right next move. If you&amp;#x27;re fitting revised terms into an existing structure, drafting the revision based on defined use cases and then tree testing with users might be the right follow-on step.&lt;/p&gt;&lt;p&gt;For this taxonomy, I&amp;#x27;ll continue with the latter of these two options. I&amp;#x27;ll discuss my process for that work and for migrating it back into my CMS in a follow-up post. If that piques your interest, stay tuned!&lt;/p&gt;</content>
  </entry>
</feed>
