skill-autoresearch — Claude Code skill: improve SKILL.md with evals

What skill-autoresearch does

The skill takes a single repo artifact — SKILL.md, an SOP, a prompt template or a workflow doc — and pushes it through a bounded improvement loop. It starts by asking whether the loop is warranted at all. The current version may already clear the bar, or the only real drift may sit in the supporting files. Without evidence of headroom, the run stops there.

When the loop does go ahead, loop-charter.md is written first. That charter pins the goal, the current baseline, the one mutable artifact, the evaluation harness, which supporting files are in scope, and the time and iteration budget. The evaluator stays untouched afterwards: a single meaningful change goes in, the same harness runs again, and both wins and reverts land in an append-only log that lives in git and survives PR review.

To start, the target needs 3-5 representative prompts or scenarios, 3-6 checks answered yes or no, and a deterministic way to inspect behavior or dry-run it. No representative prompts, no stable check — no loop. Hosted evaluation and tracing platforms such as LangSmith, Braintrust, Weave or Promptfoo sit outside this skill; requests for those get routed elsewhere.