prp-meta-skill

Gather the literal trigger phrases, the task, and gotchas first, then scaffold a directory with a SKILL.md containing YAML frontmatter (name, description) and a lean imperative body. Push bulky detail into references/ and output formats into templates/, then validate with the checklist and a real trigger test.

Classify each block as spine or extractable, move extractable content verbatim into references/ or templates/, and replace it with pointers. Always-needed output formats need a mandatory-read instruction; sometimes-needed detail gets a lazy pointer. Verify the trimmed skill drives the same process and output.

Write it in third person, state what the skill does and when to use it, and include literal user trigger phrases plus the /name invocation. Keep it under 1024 characters and avoid vague phrasing like 'helps with documents'.

References hold detail the agent reads on demand, such as schemas, edge cases, and long examples. Templates hold output shapes the result must follow; if a format is always needed, the body must instruct the agent to read it before producing output.

The most common cause is moving an always-needed output format into a lazily-loaded reference, so the agent never reads it. Pair every such extraction with a mandatory-read instruction in the body and run the behavior-preservation check.

Yes, compose specialist skills by name instead of copying their instructions into your skill. Assign each phase and artifact one owner, pass complete durable artifacts across handoffs, and avoid duplicating content between skills.