Skill Workshop

Skill Workshop

Build the smallest skill that owns the behavior, then prove its declared claims. Compose the repo's existing authoring, arena, review, and install tools; do not create another controller around them.

Non-negotiable design rules

• Write or update SPEC.md before SKILL.md or executable behavior.
• Treat a wall as evidence that the design is wrong. Re-derive the design; do not hide it behind a flag, shim, special case, fallback, or parallel path.
• Keep dry-run, static, smoke, behavioral, and representative-live evidence separate. One never stands in for another.
• Keep every attempted adapter call, failure class, and recovery action visible; omitting a failed integration attempt is itself a failed gate.
• A final review needs at least one model family other than the author. Missing capacity blocks completion; it does not authorize substitution.
• Record requested/effective effort because latency and reasoning depth trade directly against each other.
• Never persist credentials. Device config is ignored and non-secret.
• Do not add hashes, fingerprints, or timestamps as self-attestation. Actual command output and independent review own authenticity.

1. Discover the behavior

Capture the real trigger, non-trigger, user, repository, authority, and the failure or successful workflow being generalized. Describe the desired behavior without preserving accidental details of the first solution.

Redirect a narrow request to the skill named in the description instead of running the full workshop.

2. Select the fixed risk tier

Choose the highest matching tier before build. Tiers are cumulative.