| name | authoring |
| description | Use when the user asks to create, update, review, or refactor an agent skill, including requests like "turn this workflow into a skill", "make a skill for this", "fix this SKILL.md", "add references to this skill", "make this skill work in Codex and Claude", or "split this bloated skill". Creates skill directories, writes portable SKILL.md frontmatter and workflows, adds references, validates structure and token budgets. For adding executable helpers, use the tooling skill. |
| allowed-tools | Read Grep Glob Bash Write Edit Agent |
Skill Authoring
Decision Tree
- What are you doing?
- Creating a new skill from scratch -> follow "Creating a skill" below
- Improving an existing skill -> what needs work?
- Decision tree is missing or shallow -> see "Improving a decision tree" below
- SKILL.md is bloated (over ~5000 tokens) -> run
tools/token-estimate.ts <path>/SKILL.md, then extract sections over ~10 lines to references/, link from SKILL.md
- Skill instructions are too rigid or verbose -> see
references/writing-philosophy.md and rewrite: explain the why, drop heavy-handed MUSTs, remove lines that aren't pulling their weight
- Automatable tasks would benefit from a tool -> use the
tooling skill
- Missing references for detailed content -> follow "Adding a reference" below
- Frontmatter is wrong or incomplete -> see
references/skill-format.md for portable fields and product-specific extensions, then fix
- Description is vague or undertriggering -> rewrite per
references/writing-philosophy.md "Description writing"
- Not sure what's wrong -> run
tools/coverage-gap.ts <path> and tools/skill-validate.ts <path>, fix reported issues
- Adding a tool to a skill -> use the
tooling skill
- Adding a reference to a skill -> follow "Adding a reference" below
- Validating a skill -> follow "Validating" below
- Refactoring or splitting a skill -> is the SKILL.md over 500 lines or covering unrelated concerns?
- Yes, covers multiple unrelated concerns -> split into separate skill directories, one per concern
- Yes, just too long -> extract to references, don't split
- No -> the skill is fine, focus on content quality instead
- Extracting a reusable piece from a skill -> see
references/skill-taxonomy.md "Extracting skills downward". Check: could someone invoke this piece on its own and get useful output? If yes, extract it as its own lower-tier skill (atomic from workflow, workflow from agent)
- Publishing a skill to the marketplace -> use the
publishing skill
Creating a skill
1. Capture intent
Start by understanding what the skill should do. If the conversation already contains a workflow to capture (e.g., "turn this into a skill"), extract answers from it — tools used, sequence of steps, corrections made, input/output formats. Otherwise, ask:
- What should this skill enable an agent to do?
- When should this skill trigger? (what user phrases/contexts)
- What's the expected output format?
- What tier is this skill? See
references/skill-taxonomy.md — atomic (does one thing), workflow (chains atomics in a known sequence), or agent (goal-driven, decides its own path). The tier shapes the decision tree, supporting resources, validation, and composition strategy.
Proactively ask about edge cases, example files, success criteria, and dependencies. Check available MCPs for research and the codebase for similar existing skills to avoid overlap. Come prepared with context.
Agent workflow (optional)
For workflow or agent-tier skills with multiple references or unfamiliar domains, offer a research-write-validate pipeline using specialized agents. See references/agent-workflow.md for roles, flow, and the research brief format. Skip this for atomic skills or small improvements — it's overhead without quality gain. Always let the user decide.
2. Write the skill
- Pick the owning plugin and create the directory:
plugins/<plugin>/skills/<skill-name>/
- Write
SKILL.md with portable YAML frontmatter — include name and description first, then add client-specific fields only when the target client supports them (see references/skill-format.md)
- Write the decision tree — use the decision-trees skill for this
- Write conventions — the rules the agent must follow every time
- Optionally add executable helpers in
scripts/ if the skill has automatable tasks — use the tooling skill for this. Many skills work fine without any scripts. Existing skills in this repo may still use tools/; keep those paths stable unless the task is an explicit migration.
- Add references in
references/ for any detailed content over ~10 lines
Follow the writing principles in references/writing-philosophy.md — explain reasoning over rigid rules, keep the prompt lean, generalize rather than overfit to test cases.
Skill writing patterns
Keep SKILL.md under 5000 tokens (use tools/token-estimate.ts to check — line count is a rough proxy but tokens are the real constraint). If approaching this limit, add hierarchy with clear pointers to reference files. Reference files clearly from SKILL.md with guidance on when to read them. For large reference files (>300 lines), include a table of contents.
When a skill supports multiple domains/frameworks, organize by variant:
cloud-deploy/
├── SKILL.md (workflow + selection)
└── references/
├── aws.md
├── gcp.md
└── azure.md
3. Validate
- Run
tools/skill-validate.ts <path> to check structure
- Run
tools/token-estimate.ts <path>/SKILL.md to verify under 5000 tokens
Improving a decision tree
Read SKILL.md to get the skill's responsibilities, then ensure each one has a decision tree branch leading to it. Use the decision-trees skill for format and writing guidelines.
Adding a reference
- Create
references/<topic>.md — one topic per file, one level deep (no reference-to-reference chains)
- Link from SKILL.md: "See
references/<topic>.md"
- Add to the key references table in SKILL.md
Validating
- Run
tools/skill-validate.ts <path> — checks files, frontmatter, naming, line count
- Run
tools/token-estimate.ts <path>/SKILL.md — must be under 5000 tokens
- Run
tools/coverage-gap.ts <path> — compares responsibilities against content and bundled scripts/tools
Key references
| File | What it covers |
|---|
references/skill-taxonomy.md | Skill tiers (atomic, workflow, agent), choosing a tier, extracting skills downward |
references/skill-structure.md | Directory layout, SKILL.md structure, progressive disclosure |
references/skill-format.md | Portable frontmatter, Codex metadata, Claude Code extensions, naming rules, string substitutions |
references/writing-philosophy.md | Explain the why, keep prompts lean, generalize, description writing |
references/best-practices.md | Anti-patterns, keeping SKILL.md lean |
references/agent-workflow.md | Optional research-write-validate agent pipeline for complex skills |
references/advanced-features.md | Agent-specific advanced features: Codex metadata, Claude Code dynamic context, invocation control, forked execution, hooks, permissions |
Official spec vs. repo conventions
The skill format and structure references separate the portable Agent Skills baseline from Codex and Claude Code extensions. Check references/skill-format.md before adding client-specific frontmatter. The following are conventions specific to this repository — useful for organizing a skill library, but not required by the spec:
- Skill taxonomy (
references/skill-taxonomy.md) — atomic/workflow/agent tiers
- Decision trees (via the decision-trees skill) — prescribed SKILL.md structure