THE COLLECTION REVIEW / 13 SEPTEMBER 2026
A prompt is a request.
It still needs a check.
The collection is now 12 revised recipes. We retired 72 earlier templates and kept their original source history.
What changed
The earlier collection included two versions of Agent Skill Evolution Validator and two KV-cache optimisers. Other templates requested reliability scores, confidence numbers or diagnoses without a defined measurement. The earlier extraction template promised “clean JSON every time”.
Those were requests to a model, not evidence that the requested validation had happened. The current recipes define the inputs, show a worked target and say how to judge the answer. The JSON tool performs actual schema checking separately.
We retired 66 generated templates and 6 earlier templates. This is curation, not a claim that every old template failed. The retained tasks were rewritten, not certified. Specific code, arithmetic and query-plan checks are labelled where they were performed.
What an example means here
A worked example is an illustrative input and target result. It is not presented as a saved response from a particular model. Different wording can meet the checks, and a single successful result cannot establish reliable model behaviour.
Templates open as editable drafts in Workbench. They are not saved or sent automatically. The prompt bot now proposes unpublished drafts for review. Updating that source does not verify or restart the separate server.
The old routes carry retirement notices and immutable Git links. Retired templates leave search, tag listings and the sitemap. Their first recorded Git dates are shown below because the original templates did not carry publication dates.
The retired templates
72 source records · Git history dates, not invented publication dates
July 20262 templates
June 20268 templates
May 202616 templates
April 202621 templates
March 202625 templates
Inspect the source revisions, originating commits and blob hashes ↗