skill-authoringlisted

# skill-authoring A skill is loaded whole into every conversation that reaches for it, so each line either helps the agent act now or costs every future request. Two kinds of text live in a skill's repository. **Runtime** is `SKILL.md` and everything under `references/`: the only files an agent loads. **Development** is the README, the changelog, maintainer documents, script and workflow comments, vendor locks and tests: what a person reads while working on the repository. Every rule below says which side a line belongs to The agent's own instructions — `CLAUDE.md`, `AGENTS.md`, the harness's memory — are runtime under a harsher rule: loaded into every conversation rather than into the ones that reach for the topic, and with no description or trigger to earn their way in. So every rule below about runtime holds there, and the cost question is asked of each line against every request rather than every invocation [`check-skill.sh`](check-skill.sh) beside this file decides the machine-checkable half, on any skill repository: what stops a skill loading is an error, and the rest of the rules are warnings that draw attention without turning a gate red. `check-skill.sh --help` is the reference for what it checks, how a line is excused and how warnings become errors ## The boundary - **Runtime holds what an ordinary invocation needs.** Install, repository development, self-tests and checkout layout belong to the README or a maintainer document; a `Layout` section in runtime is R