In an August 14, 2026 link post titled 'Don't classify. Hallucinate!', Simon Willison described a technique from Doug Turnbull for tagging content when the set of possible tags is too large to fit into a single prompt. Willison noted that his own blog has 1,856 tags, which he said is 'likely too many to feed to an LLM in one go' and ask which apply to a given piece of content. Turnbull's suggested workaround is to have the language model generate tags freely, without giving it the existing tag vocabulary, and then use vector embeddings to compare those invented tags against the real corpus of tags to find the closest concrete matches. Turnbull's example prompt asks the model to 'create novel, never seen before' classifications for a search query, and includes examples showing the shape and hierarchy of the target tags (for instance, 'Furniture / Living Room Furniture / Coffee Tables & End Tables / Coffee Tables') to guide the model toward more useful guesses. Willison presented the method as a neat solution to the problem of matching content against a large, unwieldy tag set.
- Willison's blog has 1,856 tags, described as too many to include in a single LLM prompt
- The technique: let the model invent tags with no vocabulary, then use embeddings to find the closest real tags
- Turnbull's example prompt includes sample tag hierarchies to shape the model's output
- Posted by Simon Willison on 14 August 2026, citing Doug Turnbull
What it means for you
If you've ever tried to auto-sort a big pile of content into categories, you may have hit a wall: there are too many categories to hand the AI at once. This is a clever workaround — you ask the AI to freely guess what tags fit, then a separate step matches those guesses to your actual list. It's a practical pattern for anyone doing classification at scale, but it's a technique, not a product.
Try this
If you maintain a large tag or category list (a blog, product catalog, or help center), try the two-step approach: prompt the model to invent tags with example shapes, then use embeddings to snap those to your real vocabulary.
Who should care
Developers or technically inclined people building content tagging, product categorization, or search systems with hundreds or thousands of categories.
Skip this if
You don't work with large classification or tagging systems, or you're not comfortable with embeddings and a bit of code.
Sources: Simon Willison — read the original