Skills
Let the model hallucinate labels it has never seen, then resolve them to your real taxonomy with embeddings
Doug Turnbull's 'hypothetical classifications' technique, surfaced by Simon Willison on August 14, inverts the usual approach to large-label-space classification: instead of stuffing thousands of existing tags into the prompt, you instruct the model to invent novel labels for the item, give it a few examples only to fix the output shape, then embed both its invented labels and your real vocabulary and take nearest neighbors. Willison's own blog has 1,856 tags, well past the point where enumerating the vocabulary is practical. It is the classification-side mirror of HyDE, and it makes label-space size irrelevant to prompt cost — worth trying anywhere you currently do retrieval-then-pick over a big controlled vocabulary.
↳ Follow the thread