craft-tune

Purpose

Run an autonomous critique-and-edit loop over an existing prompt or skill until the artifact is good enough by its own critique standard — preserving the original intent throughout.

A working artifact rarely converges in one pass: diagnosis surfaces issues, edits introduce new ones, and the next round of diagnosis depends on the revised text rather than the original. craft-tune keeps cycling on its own: re-critique against the current state, apply minimal-diff edits, re-critique again. The loop terminates when the critique itself has nothing significant left to say (the Self-LGTM condition) — at which point the user gets the final artifact and the round-by-round trail of how it got there.

This is not eval-driven (no scored rubric, no eval runner — that's craft-autoresearch). It is judgment-driven: the same critique discipline that produced a trusted finding in Round 1 also decides when to stop in Round N.

Use this when

• a prompt or skill is close but needs sharpening
• a prompt "feels off" or behaves inconsistently across runs
• a working artifact needs review before being generalized or shipped
• the user wants the critique → edit rhythm to run end-to-end without hand-holding
• minimal-diff improvement across multiple rounds is better than rewriting from scratch
• setting up scored evals (the craft-autoresearch overhead) isn't worth it for this artifact