Skill Refiner / Acrazie

Run a temporary feedback campaign around one target skill. Observe how the target behaves during real use, preserve minimized evidence in an append-only journal, and promote only user-approved findings into a behavior record. Use an ADR only for durable, nontrivial trade-offs, not for every observation or stylistic preference.

Do not patch, rewrite, or otherwise modify the target SKILL.md. This skill documents what should be preserved or changed; a separate editing workflow may implement those decisions later.

Campaign authority and loading

Refiner remains explicit-only. Its activation identifies a requested campaign, not permission to edit the target, execute its tasks or install dependencies. An explicit project kit or stricter instruction also gates the target and every other dependency: propose from metadata and wait for the user's explicit command before reading a body or dispatching it. Re-reading a previously activated target for digest verification is within the approved campaign boundary; a new target or returning execution specialist follows its applicable renewed loading gate. Never activate a chain or bypass it through a subagent. These are instruction boundaries, not filesystem controls or live host proof.