prp-meta-skill

# PRP Meta-Skill — Author & Refactor Skills A PRP-style runbook for two jobs: - **Create** a new skill from scratch (or from an existing command/prompt). - **Refactor** an existing skill — split detail into `references/`, move output formats into `templates/`, and trim `SKILL.md` to a lean spine of pointers. This skill is built the way it teaches: a lean body that defers detail to `references/`. Follow that example. ## Prescribe the craft, not the content This skill is opinionated about **how to build a skill well** and deliberately agnostic about **what any given skill — or the artifact it produces — should contain**. - **Prescribe (firm, universal):** progressive disclosure, third-person trigger-rich descriptions, imperative body, no-duplication, lean body, every reference wired, deliberate invocation control, validation that fits. See `references/skill-standards.md`. - **Do NOT prescribe (per-project, the author's call):** the sections a plan / PRD / report should contain, what a project's template looks like, which phases exist, the domain vocabulary. There is no canonical output shape — guide the author to a good decision, never hand them a fixed one. Restrictions are not rigidity: be strict on the craft so the author stays free on the content. ## The PRP lens — applied to the skill *type* Classify the skill first (workflow / artifact-generator / knowledge-reference / tool-wrapper — see `references/skill-standards.md` → Skill types), then apply only the principles that fit: 1. **Context is King** — give the agent ALL the context it needs (patterns, gotchas, schemas, examples) via whichever source fits: inline, bundled and disclosed on demand, pointed to by file path or URL, or gathered from the user at runtime. Curate it — don't dump it. (`references/skill-standards.md` → Context sources.) 2. **Validation that fits** — *workflow* skills ship verifiable gates, and prefer an external, authoritative check (exit code, file presence) over the agent's own "done" sentinel. A knowledge/reference skill has nothing to validate — don't bolt a loop onto it. 3. **Information dense** — real trigger phrases, real examples, real `file:line`. No filler, no restating what the model already knows. 4. **Progressive success** — ship a minimal SKILL.md that triggers correctly first, validate, then enrich with references. Don't build all the references before the spine works. ## Step 0 — Pick the mode - **Creating** a new skill → follow `references/creating-skills.md`. - **Refactoring / trimming** an existing skill → follow `references/refactoring-skills.md`. - Both modes obey the same rules (frontmatter spec, writing style, progressive disclosure, no-duplication, invocation control) → read `references/skill-standards.md` first. If the argument is blank, ask which mode and what the skill/target is. Do not guess. ## Create — quick spine (full detail in references/creating-skills.md) 1. **Gather context (PRP):** collect the concrete phrases that should trigger the skill, the task it performs, the gotchas, and existing patterns to mirror. 2. **Plan resources:** what repeats → a `scripts/` script; what informs thinking → a `references/` doc; what is reused in the output → a `templates/`/`assets/` file. 3. **Scaffold:** copy `templates/SKILL.template.md` into `.claude/skills/<name>/SKILL.md`; create `references/`, `templates/` only as needed. 4. **Write the spine:** third-person trigger-rich `description`; imperative, lean body; push detail to references. 5. **Validate & iterate** (`references/validation.md`): checklist → skill-reviewer → trigger test. ## Refactor — quick spine (full detail in references/refactoring-skills.md) 1. **Inventory** the target SKILL.md. Classify each block: _spine_ (keep) vs _extractable_ (output-format templates, schemas, long worked examples, exhaustive pattern catalogs, troubleshooting, edge-case lists). 2. **Extract verbatim** into the TARGET skill's own `references/` (or `templates/` for output formats). Do not reword anything that affects behavior. 3. **Replace with a pointer.** For **always-needed** content (e.g. an output format the agent must always follow), add a _mandatory-read_ instruction: "Before producing output, read `templates/<x>.md`." For **sometimes-needed** content, a lazy pointer ("For edge cases, see `references/<x>.md`"). 4. **Behavior-preservation check** — the trimmed skill + references must drive the SAME process and SAME output as before. Nothing lost, nothing duplicated. 5. **Validate** (`references/validation.md`). > The single biggest refactor risk: moving an _always-needed_ output format into a lazily-loaded reference, so the agent forgets to read it and the output silently changes. Always pair such extractions with a mandatory-read instruction. See `references/refactoring-skills.md`. ## Resources - `references/skill-standards.md` — frontmatter spec, progressive disclosure, writing style, no-duplication, invocation control, cross-provider notes - `references/creating-skills.md` — full PRP-style create runbook - `references/refactoring-skills.md` — full split-and-trim runbook with a before/after example - `references/validation.md` — validation gates, checklist, skill-reviewer, trigger test - `templates/SKILL.template.md` — lean SKILL.md skeleton - `templates/reference.template.md` — skeleton for an extracted reference/template file