Skill creator

Skill Creator

<instructions>

Creates production-quality skills that follow the Anthropic skill spec. A skill is a RECIPE that changes Claude's behavior — not documentation, not a prompt.

5-Minute Checklist

Before diving into the full workflow, verify these 5 essentials:

1. Does this skill change Claude's behavior? (If removing it changes nothing, don't build it)
2. Is the description pushy enough? (30+ trigger keywords, ends with "even if...")
3. Is the body under 2,000 words? (Deep content goes in references/)
4. Does the workflow use imperative form? ("Run X" not "You should run X")
5. Is there at least one concrete example? (Claude follows examples > instructions)

When NOT to Create a Skill

• The behavior already exists natively in Claude Code (most prompt wrappers are unnecessary)
• The "skill" is just documentation — if it doesn't change workflow steps, it's a reference file
• Another skill already covers this domain — extend it instead of creating overlap
• The task is a one-off — skills are for repeated patterns, not single-use instructions

Workflow

1. Capture intent. Ask: what task does this skill handle? What triggers it?

1b. Web research (MANDATORY for domain skills). Before writing any skill content:

• WebSearch for "{domain} best practices {current year}" — get current patterns, not stale training data
• WebSearch for "{domain} common mistakes" — build the anti-patterns section from real failures
• WebFetch the official docs for any API/library the skill covers — get correct method names, parameters, versions
• Include findings in the skill body and references/ — the skill must reflect CURRENT reality, not training-data memory
• Why: Claude hallucinates outdated APIs. A skill built on stale knowledge is worse than no skill.

2. Generate the directory:

   bash .claude/skills/skill-creator/scripts/scaffold.sh SKILL_NAME
   

3. Write the frontmatter. Follow the pushy description formula:

   description =
     WHAT it does (1 sentence)
     + ACTIONS that trigger it (10+ verbs)
     + OBJECTS it applies to (10+ nouns)
     + COMMANDS that invoke it
     + CONTEXTS where it fires without explicit request
     + "Even if the user doesn't explicitly mention X, use this when Y"
   

   Target 30+ trigger keywords. Claude undertriggers modest descriptions.

4. Write the body. Under 2,000 words. Imperative form.

   • "Run the script" not "You should run the script"
   • "Always validate" not "Don't skip validation"
   • Include one concrete example inline
   • Move deep content to references/

5. Add scripts/ if the skill involves deterministic work:

   • Scanning files for patterns
   • Detecting frameworks or tools
   • Formatting structured output
   • Do NOT put LLM reasoning in scripts

6. Add examples/ with concrete input → output samples.

7. Validate. Read references/quality-checklist.md and verify all items pass.

Rules

• SKILL.md body: 1,500-2,000 words maximum
• Description: 30+ trigger keywords minimum
• Scripts: deterministic only, output structured text
• Scripts: SKILL.md says "Do NOT read the script file, just execute it"
• Every reference file mentioned in SKILL.md with load condition
• Removing the skill must change Claude's behavior (not just add info)
• disable-model-invocation: true ONLY for dangerous skills

References

For the complete skill engineering guide: references/skill-engineering-guide.md For the quality checklist: references/quality-checklist.md For a sample skill output: examples/sample-skill.md

</instructions>