Skill creator

Create and audit skills through one cross-runtime quality contract: precise triggering, appropriate freedom, safe authority, progressive disclosure, and evidence proportional to maturity and risk.

Write machine-facing skill prose in English for token efficiency. Keep literal user triggers, UI labels, domain terms, and quotations in their native language where the exact text matters.

Select action and evidence

Choose the action first:

• Create: the target skill does not exist or the user requests a new one.
• Audit: inspect an existing skill, read the complete skill and reachable resources, validate structure, and report findings without changing files.
• Fix: apply explicitly authorized improvements to an existing skill, then recheck the changed contract.

Choose the evidence tier separately:

• Static (default): inspect files, run structural validation, and judge contracts and dependencies without live model invocations.
• Behavioral (optional): run clean-context scenarios from references/evals.md when the user requests live evaluation or a stable/shared/high-risk skill needs behavioral proof for a load-bearing claim.