[Demo] How to re-categorize content at scale using LLMs
Summary
Large Language Models (LLMs) are to language as spreadsheets are to numbers: tools for modeling, exploration, and development. Among their many capabilities, LLMs can alleviate chores related to the design and implementation of information architectures. But doing so requires venturing beyond chat-based interfaces. In this brief demonstration, we'll see how to use OpenAI's API and a few open source command line tools to re-categorize content in a 1,000+ page website. The techniques demonstrated can be extended to other common content organization tasks.
Key Insights
-
•
Manual retagging of 1,200 blog posts would take about 10 hours, but leveraging GPT-4 reduced active human time to about 2 hours.
-
•
Using GPT-4 via command line and shell scripts enables automated tagging outside typical chat interfaces.
-
•
An organically grown taxonomy over 20 years contained unclear acronyms and inconsistent tag forms that GPT initially struggled with.
-
•
Cleaning and standardizing the taxonomy before prompting GPT is critical for effective AI assistance.
-
•
A review step of AI-suggested tags in CSV format allows human correction to avoid hallucinations entering production.
-
•
GPT-4 can propose new and useful tags outside the original taxonomy, enriching content classification.
-
•
The four-step GRU framework (Gather, Review, Update, Wrap up) balances automation with human oversight.
-
•
Storing blog content as markdown files simplifies integrating AI workflows via scripting and file manipulation.
-
•
The approach is adaptable and scalable to other CMS platforms by replacing scripting with API calls.
-
•
Taxonomies should use clear, unambiguous terms to improve both human and AI understanding.
Notable Quotes
"Some of the older content has discoverability problems, which is typical with blogs."
"Doing this tagging manually would have taken me around 10 hours of mind-numbing work."
"I’m actually using GPT-4, but not via the chat interface—I'm calling it from the Mac’s command line."
"I had to clean the taxonomy up because GPT wouldn’t know what to do with acronyms like TAOI."
"I save the proposed tags to a CSV file so I can preview and edit them before applying the changes."
"A middle review step prevents hallucinations from making it into the production site."
"GPT-4 functioned as an assistant not just in retagging but also in improving the taxonomy itself."
"The entire process took about three hours from start to finish, about a fifth of the manual time."
"Use clear and obvious terms in taxonomies—unusual acronyms won’t make sense to GPT or others."
"You need to review proposed changes before committing them to production, otherwise errors sneak in."
Or choose a question:
More Videos
"We look to work with authors we like and enjoy being around because writing a book is a long, collaborative journey."
Louis RosenfeldCoffee with Lou: Should You Write a (UX) Book?
March 7, 2024
"Legal focuses on risk mitigation; research focuses on ethical best practices."
Theresa MarwahHow Atlassian is Operationalizing Respect in Research
February 27, 2020
"Design system decision records prevent knowledge being stuck in one person’s head."
Dan DonaldDesign Systems as a Vehicle for Systemic Change
June 1, 2023
"Mobile accessibility features are built in and ready to go without needing to install or configure anything extra."
Sam ProulxMobile Accessibility and You
June 9, 2022
"Teams are better at solving problems and coming up with innovative ideas than solo practitioners."
Megan BlockerGetting to the “So What?”: How Management Consulting Practices Can Transform Your Approach to Research
March 26, 2024
"Cloud code wasn’t programmed to do this; it’s a very lightweight system where the model does most of the work."
Peter Van DijckHands on AI #3: Claude Code for UX people
October 22, 2025
"Think of AI more like a junior associate, a really powerful one, not a replacement."
Louis Rosenfeld Billy Carlson Jon Fukuda Maria TaylorHow AI will Change DesignOps Tooling
October 3, 2023
"Breaking enterprise into small segments lets you apply research methods effectively at scale."
Steve Portigal Chris Chapo Kelly Goto Christian RohrerDiscussion
May 13, 2015
"The impact of design ops isn’t as easy to define or predict like other business value measures."
Frances YllanaTheme 2 Intro
September 24, 2024