writing-skills

---

name: writing-skills description: Use when creating a new Claude Code skill from scratch, extending an existing skill, or reviewing a skill for structure compliance. Covers anatomy, frontmatter fields, progressive disclosure, 500-line limit, and the 6-step creation checklist. Triggers: "create a skill", "write a skill", "add a skill", "new skill for X", "skill structure", "how do I write a skill". allowed-tools: Read, Write, Edit, Glob, Grep metadata: triggers: create a skill, write a skill, add a skill, new skill, skill structure, skill authoring, skill template, SKILL.md related-skills: subagent-driven-development, plan-mode-review, documentation-generation domain: workflow role: specialist scope: design output-format: document last-reviewed: "2026-03-16"

Writing Skills

Iron Law

NO SKILL AUTHORING WITHOUT READING AN EXISTING SKILL FIRST

Read at least one existing SKILL.md in .claude/skills/ before writing your own. Understand what the pattern looks like in practice, not just in spec.

Anatomy of a Skill

.claude/skills/<skill-name>/
  SKILL.md                    <- Body: <=500 lines. Loaded on trigger.
  references/                 <- Deep content: loaded as needed
    <topic>.md
  assets/                     <- Diagrams, templates, static files
  scripts/                    <- Executable helper scripts

Frontmatter fields (see references/skill-anatomy.md for full spec):

---
name: my-skill                          # Required -- matches directory name
description: [trigger text]            # Required -- this is how the skill gets selected
allowed-tools: Read, Write, Glob        # Strongly recommended -- limits available tools
last-reviewed: "YYYY-MM-DD"            # Recommended -- for /audit-skills staleness check
metadata:                              # Optional -- arbitrary key/value
  version: "1.0"
---

Note on allowed-tools: This onboarding kit uses allowed-tools in 24 of 29 skills. It IS valid per the official spec. Always include it -- it makes tool availability explicit and prevents the skill from inheriting an unrestricted toolset.

Progressive Disclosure (3 Levels)

Level 1 -- Always loaded:    Frontmatter (name, description, allowed-tools)
Level 2 -- Loaded on trigger: SKILL.md body (<=500 lines)
Level 3 -- Loaded as needed:  references/ files (deep content, patterns, templates)

Rule: If content is referenced once per session, put it in the body. If it's deep reference material consulted occasionally, put it in references/.

The 500-Line Limit

SKILL.md body must stay under 500 lines. When approaching the limit:

1. Move detailed examples to references/
2. Move stack-specific patterns to references/<stack>.md
3. Keep only: Iron Laws, decision trees, quick checklists, cross-references in the body

What NOT to Include

Do NOT create these files in a skill directory:

• README.md -- skills are not packages, no installation docs needed
• INSTALLATION_GUIDE.md -- same reason
• CHANGELOG.md -- version history belongs in git, not skill files
• CONTRIBUTING.md -- meta-meta documentation that nobody reads

Description as the Trigger Mechanism

The description field is how Claude selects this skill. Write it like a search-optimized trigger:

# BAD -- too vague
description: Skill for building things with Flutter

# GOOD -- includes all trigger words, usage contexts, explicit "when to use"
description: Use when building Flutter screens, Riverpod providers, Freezed models,
  or widget tests. Covers clean architecture, Firebase integration, and adaptive UI.
  Triggers: "Flutter", "Dart", "mobile app", "Riverpod provider", "widget test".

Rule: Include every word a developer would type when they need this skill.

Minimal Correct Skill Example

---
name: my-linter
description: Use when running or configuring the project linter. Covers ESLint,
  Prettier, and auto-fix workflows. Triggers: "lint", "fix lint", "eslint config".
allowed-tools: Bash, Read, Edit
last-reviewed: "2026-01-01"
---

**Iron Law:** Always run the linter before reporting a file clean.

# My Linter Skill

## Verify

```bash
npx eslint . --max-warnings 0   # must exit 0

The `description` gets the skill selected; the Iron Law prevents silent skips; the verify command gives a binary pass/fail gate.

## 6-Step Creation Checklist

[] 1. NAME -- matches directory, lowercase-hyphenated, specific domain [] 2. DESCRIPTION -- includes all trigger words; passes the CSO test above [] 3. ALLOWED-TOOLS -- listed explicitly; minimum needed for the skill's tasks [] 4. BODY -- Iron Law first; decision trees > prose; <=500 lines [] 5. REFERENCES -- deep patterns offloaded; each reference file < 300 lines [] 6. TEST -- load the skill, ask it a question, verify it triggers correctly

## Verify Command (required in every skill)

Every skill MUST include a "Run X to verify" gate — a command or checklist that produces a binary pass/fail result. Place it in its own `## Verify` section near the bottom of SKILL.md.

```markdown
## Verify

Run:  <command>   # exit 0 = green
-or-
Checklist:
- [ ] <criterion 1>
- [ ] <criterion 2>

Cross-Stack Skill Pattern

For skills that cover multiple tech stacks (Java, NestJS, Python, Flutter):

## Stack Dispatch

| Stack   | Reference File |
|---------|---------------|
| Java    | references/patterns-java.md |
| NestJS  | references/patterns-nestjs.md |
| Python  | references/patterns-python.md |
| Flutter | references/patterns-flutter.md |

Keep the body stack-agnostic. Push all stack-specific code into reference files.

References

• references/skill-anatomy.md -- full frontmatter field spec, directory conventions
• references/skill-authoring-spec.md -- condensed official spec, what NOT to include, CSO rules