# Skill Writing Guide

Best practices for writing effective Claude Code skills.

## Two Categories of Skills

1. **Capability uplift** — teaches the agent something it couldn't do before (scaffold component, run audit, deploy)
2. **Encoded preference** — captures your specific way of doing something the agent could already do (commit style, review checklist, naming conventions)

Know which you're building — it changes how much detail to include.

## Description Optimization

The description is the most important line. It determines when the skill gets triggered.

- List trigger contexts explicitly: "Use when the user wants to X, Y, or Z"
- Think about should-trigger / should-not-trigger scenarios
- A slightly "pushy" description is better than a vague one
- Test: would this description make the model select this skill for the right prompts?

## Writing Instructions

### Explain WHY, not just rules
- Bad: "MUST use semantic HTML"
- Good: "Use semantic HTML elements (nav, main, aside) because screen readers depend on landmarks for navigation"

### Avoid heavy-handed MUSTs
- Reserve MUST/NEVER for genuine constraints (security, data loss)
- For preferences, explain the reasoning and let the agent make good decisions

### Progressive disclosure
Three levels of instruction loading:
1. **Frontmatter** — always loaded (name, description). Keep minimal.
2. **Body** — loaded when skill is invoked. Core instructions here.
3. **Bundled resources** — loaded on demand via `Read`. Put reference tables, specs, examples here.

Use bundled resources (`references/`, `scripts/`, `assets/`) for content that would bloat the main SKILL.md.

### Every sentence should change behavior
- Delete filler: "It is important to...", "Make sure to...", "Please note that..."
- Delete obvious instructions the agent would do anyway
- Test: if you removed this sentence, would the output change? No → delete it.

## Structure Conventions

### Project conventions (this repo)
- Always set `disable-model-invocation: true`
- Use H1 for the skill title (short action phrase)
- Reference `$ARGUMENTS` early in the body
- Use `!` backtick for live data injection (git diff, file listings)
- Numbered steps, imperative voice
- Output format in a fenced markdown block if structured

### Bundled resources pattern
```
.claude/skills/my-skill/
  SKILL.md              # Main instructions
  references/           # Specs, guides, schemas
  scripts/              # Shell scripts, templates
  assets/               # Static files
```

Reference from SKILL.md: `Read ${CLAUDE_SKILL_DIR}/references/spec.md`

## Length Guidelines

- Simple skills (encoded preference): 30-50 lines
- Standard skills (capability uplift): 50-100 lines
- Complex skills (multi-mode, research): 100-200 lines
- Maximum: 500 lines (if exceeding, split into bundled resources)

## Common Mistakes

1. **Overfitting to test cases** — write general instructions, not scripts for specific inputs
2. **Too many rules** — the agent ignores rules after ~20 constraints. Prioritize.
3. **No examples** — for complex output formats, show one complete example
4. **Ignoring conversation context** — skills without fork can use prior conversation. Leverage it.
5. **Forgetting edge cases** — what happens with empty input? Invalid arguments? Missing files?

## Improvement Workflow

1. Draft the skill
2. Test with 3-5 realistic prompts
3. Review output — does every instruction change behavior?
4. Remove filler, tighten descriptions
5. Add edge case handling for failures observed in testing
6. Re-test after changes

## Evaluation Criteria

When reviewing a skill, score against:
- **Trigger accuracy** — does the description match the right prompts?
- **Instruction clarity** — can the agent follow without ambiguity?
- **Output quality** — does the skill produce useful, consistent results?
- **Conciseness** — is every line earning its place?
- **Robustness** — does it handle edge cases and errors?