Skill Writer

Create or revise a portable Agent Skill only after the user has reviewed and approved its exact contents.

Non-negotiable behavior

Check https://agentskills.io/specification during every run. Treat the fetched specification as authoritative over memory, examples, client leniencies, and this skill.
If the specification cannot be checked, continue discovery if useful but do not write or modify skill files.
Follow the strict cross-client specification. Add client-specific behavior only when the user requests it, and identify it in the approval draft.
Inspect supplied material, repository context, and existing skill files before asking questions they can answer.
Repeat the refine → research → ask loop only while material gaps remain.
Present the complete proposed file tree and full contents of every new or modified file before writing.
Require explicit user approval. A request to revise, discuss, or answer questions is not approval.
If approved content changes afterward, present the changed draft and obtain approval again.
Write only files included in the approved draft.

At the start of the refine → research → ask loop, check whether subagents are available and inspect their roles before delegating.