Skip to main content
Each cookbook starts from a real dataset and a question someone would want answered about it. It shows the spec, what was run, what came out, and where the approach fell short. Every number comes from a run in the repository: the example specs are in prototype/examples/, and the design notes behind them in docs/04-design.md.

Grade Claude Code sessions

Did each turn work? About 90% right against reviewers, and diff caught a wording change that made it worse.

Guard a coding agent's commands

Stops 11.6% of 1,315 real commands and lets through an estimated 1% (0.2–6%) that needed a person.

Route banking intents

77 intents. The answer key turned out to be wrong more often than the model.

Predict whether a patch passes

A confident “no” is reliable; a confident “yes” is not.

100,000 product reviews

57.5 requests per second, $1.55, and a cached re-run in 4 seconds.
The same specs on a second engine are compared in Engines; automatic rewrites with hunch suggest are in Change a spec.

​
Not yet

  • No one outside the project has used hunch.
  • The review panels were AI reviewers working from each row alone. The own-session reviews were by the developer who wrote the sessions. Neither is an independent human panel.
  • suggest has not produced a rewrite that beat hand-written option descriptions.
  • The server has only run locally.