Skills Create — the meta-skill
A skill that authors and refactors skills. Two jobs:
- Create a new skill from scratch (or from an existing prompt/command).
- Refactor a fat skill — split detail into
references/, move output shapes intotemplates/, trimSKILL.mddown to a lean spine of pointers.
It's built the way it teaches: a lean body that defers detail to references/. That's progressive
disclosure, and this skill is its own worked example. It's all composable markdown: a skill is a
SKILL.md plus optional files the agent loads only when it reaches for them.
Prescribe the craft, not the content
Be strict on how a skill is built and agnostic on what any given skill — or its output — should contain.
- Prescribe (firm, universal): progressive disclosure · a third-person, trigger-rich
description· an imperative, lean body · no duplication · every bundled file wired · deliberate invocation control · validation that fits the skill type. (Seereferences/skill-standards.md.) - Do NOT prescribe (the author's call): the sections a plan/PRD/report should have, the domain vocabulary, the output shape, which phases exist. There is no canonical output — guide the author to a good decision, never hand them a fixed one.
Strict on the craft so the author stays free on the content. (Guide, don't prescribe — and take-vs-build: you build a skill when you own the process and want it to follow YOUR way.)
Classify the skill first
Pin the type before applying the craft — the guidance is proportional, not one-size-fits-all:
| Type | What it is | Apply |
|---|---|---|
| Workflow | a multi-step procedure (plan, review, ship) | the full lens incl. verifiable validation gates |
| Artifact-generator | produces a document/output | Context-is-King + a suggested (never mandated) output shape |
| Knowledge / reference | facts the agent consults | Context-is-King + information-dense; no phases, no validation loop |
| Tool-wrapper | drives a script / CLI / API | a deterministic script + sharp triggers; validation = the tool's own exit code |
A skill can blend types — apply the union of what fits. Never force a workflow's machinery (phases, loops, output
skeletons) onto a knowledge skill. Full detail: references/skill-standards.md → Skill types.
Step 0 — pick the mode (from $ARGUMENTS)
Parse $ARGUMENTS to pick the mode and the target:
- Starts with
create [<name>](or is clearly a new-skill ask) → create mode; use<name>as the skill name if one was given → followreferences/creating-skills.md. - Starts with
refactor <path/to/SKILL.md>(or points at an existing skill) → refactor mode on that path → followreferences/refactoring-skills.md. - Blank or unclear → ask which mode and what the skill/target is. Don't guess.
Both modes obey the same craft rules → read references/skill-standards.md first.
Create — quick spine (full detail: references/creating-skills.md)
- Gather context — the literal phrases that should trigger it, the task start-to-finish, the gotchas, the patterns to mirror. Ask the user; don't write yet.
- Plan the resources — what repeats →
scripts/; what informs the work →references/; what shapes the output →templates/; what you don't own / changes upstream → cite a path/URL; what only exists at runtime → gather it (ask the user, read git/codebase). - Scaffold — copy
templates/SKILL.template.mdto.claude/skills/<name>/SKILL.md; addreferences//templates/only as the plan needs. - Write the spine first — third-person trigger-rich
description; imperative, lean body; push detail to references. Get it triggering before you write the references. - Validate & iterate —
references/validation.md(checklist → trigger test → run it for real).
Refactor — quick spine (full detail: references/refactoring-skills.md)
- Inventory the SKILL.md — mark each block spine (keep) vs extractable (output templates, schemas, long examples, exhaustive pattern/edge-case lists).
- Extract verbatim into the target skill's
references/(ortemplates/for output shapes) — don't reword anything that affects behavior. - Replace with a pointer — for always-needed content, a mandatory-read line ("Before producing output,
read
templates/<x>.md"); for sometimes-needed content, a lazy pointer. - Behavior-preservation check — same process, same output as before. Nothing lost, nothing duplicated.
- Validate —
references/validation.md.
The #1 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 an extraction with a mandatory-read line.
When to build a skill at all
Build one when you own a process and want the agent to follow your way of it, repeatedly — the take-vs-build rule. A rough threshold: you've prompted the same thing ~3 times (the Rule of Three) → bank it as a skill. Don't build a skill for a one-off, or for a tool whose owner already ships a good one.
Resources
references/skill-standards.md— the craft rules: skill types, anatomy, context sources, frontmatter spec, progressive disclosure, writing voice, no-duplication, "structure implies a maintainer", wiring, portability.references/creating-skills.md— the full create runbook (incl. porting an existing prompt/command).references/refactoring-skills.md— the full split-and-trim runbook, with a before/after.references/validation.md— the validation gates (structure · description · body · disclosure · behavior · trigger test).templates/SKILL.template.md— the lean SKILL.md skeleton to scaffold from.
