Evals
2 promptsPrompts with test cases.
The prompts my workbench runs are files with versions. Each one has test cases, and the same cases run on several models.
How this works
Every prompt is a file in the repo, with one file per version and a note on what changed. A test case is an input and a list of checks on the answer, for example that it is valid JSON, that it holds a certain word, or that a second model agrees with a statement about it. A case passes when all its checks pass. A model's score is the average share of checks passed over all cases. A run sends every case to every model once, through Cloudflare Workers AI, and stores the scores. The page shows the latest run of each prompt version.
Session to lesson
2 versions, 8 casesRead the notes on a coding session and answer with JSON, a title, what was built, what went wrong, what I would do differently, and tags.
No runs yet.
Summarise a saved link
2 versions, 8 casesRead the text of one web page and answer with JSON, a title, a summary of at most 120 words, exactly three points and up to five tags.
No runs yet.