/skill-creator

사용법

Claude Code CLI 또는 연결된 에이전트에서 호출합니다.

상세 설명

The Skill Creator is Anthropic's tool for developing, testing, and refining Claude skills through an iterative workflow.

The fundamental loop:

1. **Capturing Intent** — Understand what the skill should do, when it triggers, expected outputs, and whether test cases are needed
2. **Drafting** — Write the SKILL.md with clear instructions and examples
3. **Testing** — Run 2-3 realistic test prompts with and without the skill
4. **Evaluation** — Use the eval viewer to let users review outputs and provide feedback
5. **Iteration** — Improve the skill based on feedback and repeat

```
~/.claude/skills/{skill-name}/
├── SKILL.md              ← Required. Core instructions. <500 lines.
├── references/            ← Optional. Detailed docs.
│   └── {topic}.md
├── evals/                 ← Optional. Test cases.
│   └── evals.json
└── scripts/               ← Optional. Helper scripts.
```

### SKILL.md Frontmatter

```yaml
---
name: my-skill              # kebab-case, max 64 chars. Becomes /slash-command.
description: >               # CRITICAL: Claude uses this to decide when to invoke.
  What this skill does.       # Max ~200 chars effective.
  Invoke when: trigger keywords.
---
```

**Optional frontmatter fields:**

| Field | Default | Description |
|-------|---------|-------------|
| `disable-model-invocation` | false | true = manual `/skill` only |
| `user-invocable` |

### Progressive Disclosure
Skills load metadata first (~100 words), then SKILL.md body (<500 lines), with bundled resources as needed. Keep SKILL.md lean; move details to `references/`.

### Generalization Over Overfitting
When iterating, avoid changes tailored only to test examples. Branch out with different metaphors and patterns.

### Explanation > Rigid Rules
Rather than all-caps directives like "ALWAYS", explain *why* something matters. Models apply reasoning beyond rote instructions.

### Lean Prompts
Remove instructions that aren't productive. Watch transcripts for wasted effort.

### Step 1: Capture Intent

Ask:
- What does this skill do?
- When should it trigger? (keywords)
- What tools does it need?
- What does good output look like?
- Are there edge cases?

### Step 2: Write SKILL.md

```markdown
---
name: example-skill
description: >
  Does X when Y. Invoke when: "keyword1", "keyword2", "keyword3".
---

# Skill Title

Brief description of what this skill does.

Step-by-step instructions Claude follows.

Concrete examples of input → output.

Common mistakes to avoid.
```

### Step 3: Quality Checklist

- [ ] `name` matches directory name (kebab-case)
- [ ] `description` has 3+ trigger keywords
- [ ] `description` includes "Invoke when:" pattern
- [ ] Body ≤ 500 lines
- [ ] Step-by-step instructions present
- [ ] Next skill call noted (if chained)
- [ ] "What NOT to do" section present
- [ ] Code/Cypher examples are copy-paste ready

1. Save test cases to `evals/evals.json` (no assertions initially)
2. Spawn with-skill and baseline runs in parallel
3. Draft quantitative assertions explaining what each checks
4. Use `eval-viewer/generate_review.py` to show results before iterating

After the skill is complete, optimize its description for triggering accuracy:

1. Create 20 eval queries — 10 that should trigger, 10 that should NOT
2. Run the automated loop with `scripts/run_loop`
3. Iterate description until trigger accuracy is high

### Linear Chain (APT example)
```
/apt → /apt-sa → /apt-sp → /apt-st → /apt-scw
```

Each skill's last section should point to the next:
```markdown

SP decomposition complete. Run `/apt-st` to define Contracts.
```

### Conditional Routing
```
/apt → Phase Detection
  ├── PH3 → /apt-sp
  ├── PH4 → /apt-st
  └── PH5 → /apt-scw
```

### Arguments
```markdown
Target: $ARGUMENTS
```
Usage: `/my-skill some-argument`

### Dynamic Context (preprocessed)
```markdown
Current status: !`git status`
Recent log: !`git log --oneline -5`
```
`!` backtick executes before skill loads — result injected into prompt.

| Anti-pattern | Fix |
|-------------|-----|
| SKILL.md > 500 lines | Move to references/ |
| Vague description | Add "Invoke when:" + 3 keywords |
| Hardcoded paths | Use $ARGUMENTS or config |
| No next-step guidance | Add "Next Step" section |
| No testing | Create evals/evals.json |
| ALL-CAPS rules | Explain why instead |

메타데이터

같은 카테고리