Anthropic Skill Creator

Anthropic-Style Skill Creator

Workflow for creating and improving Cursor skills. Adapted from anthropics/skills skill-creator.

Core loop: Capture intent → Draft skill → Run 2–3 test prompts → User reviews outputs → Improve from feedback → Repeat until satisfied.

1. Capture Intent

Before writing, clarify:

1. What should this skill enable the agent to do?
2. When should it trigger? (user phrases, file types, context)
3. Expected output format? (templates, structure)
4. Test cases? Skills with verifiable outputs (file transforms, data extraction, code) benefit from test prompts. Subjective outputs (tone, design) often don’t — suggest based on type, let user decide.

If the conversation already shows a workflow (“turn this into a skill”), extract from history: steps, tools, corrections, formats. Fill gaps with the user and confirm before drafting.

2. Interview and Research

Ask about edge cases, I/O formats, example files, success criteria, dependencies. Don’t write test prompts until this is clear.

If useful, search the codebase or web for similar skills, conventions, or docs — come with context so the user doesn’t have to repeat.

3. Write the SKILL.md

Paths in Cursor:

Do not create skills in ~/.cursor/skills-cursor/ — that’s for Cursor’s built-in skills.

Frontmatter:

• name: Lowercase, hyphens, unique id.
• description: Primary trigger. Include both what the skill does and when to use it (phrases, contexts). Cursor tends to undertrigger — be a bit “pushy”, e.g. “Use this skill whenever the user mentions dashboards, data viz, or internal metrics, even if they don’t say ‘dashboard’.”
• compatibility: Optional; only if specific tools/deps are required.

Skill anatomy

skill-name/
├── SKILL.md       # Required — frontmatter + instructions
├── references/    # Optional — docs to read when needed
├── scripts/       # Optional — runnable helpers
└── assets/        # Optional — templates, static files

Progressive disclosure

• Metadata (name + description): Always in context.
• SKILL.md body: When skill triggers; keep under ~500 lines.
• Bundled files: Load only when needed; link from SKILL.md with “when to read”.

Patterns:

• Keep SKILL.md under 500 lines; if close, add hierarchy and pointers to reference files.
• For long references (>300 lines), add a short ToC in SKILL.md.
• Multi-domain skills: one SKILL.md + references/aws.md, references/gcp.md, etc.; instruct to read only the relevant one.

Writing style

• Prefer imperative instructions.
• Explain why instead of stacking MUSTs. Skills that teach reasoning generalize better.
• Avoid overfitting to one example; keep instructions general.

Output format:

## Report structure
Use this template:
# [Title]
## Executive summary
## Key findings
## Recommendations

Examples:

## Commit format
**Example 1:**
Input: Added user auth with JWT
Output: feat(auth): implement JWT-based authentication

Principle of lack of surprise

No malware, exploits, or misleading intent. Don’t create skills for unauthorized access or data exfiltration. Roleplay/creative skills are fine if described honestly.

4. Test cases

After the draft, propose 2–3 realistic test prompts — what a real user would say. Confirm with the user, then run them.

In Cursor: You have the skill in context. For each prompt, follow the skill’s instructions and produce the output. Save outputs under a workspace folder, e.g. skill-workspace/iteration-1/eval-1/, and show the user (files, snippets, or summary).

Optional: record prompts in evals/evals.json for later iterations:

{
  "skill_name": "example-skill",
  "evals": [
    {
      "id": 1,
      "prompt": "User's task prompt",
      "expected_output": "What good looks like",
      "files": []
    }
  ]
}

5. Review and iterate

• Show the user each test case: prompt + output.
• Ask for feedback (inline or structured).
• Improve the skill from feedback:
  • Generalize — fix the underlying pattern, not only the example.
  • Trim — remove instructions that don’t help; check if the agent is wasting steps and simplify.
  • Explain why — turn rigid MUSTs into reasoning so the agent can generalize.
  • Reuse repeated work — if every run builds the same helper, put it in scripts/ and tell the skill to use it.

Then: run the same (or updated) test prompts again into a new iteration folder, get feedback again, repeat until the user is satisfied or feedback is empty.

6. Description tuning (optional)

After the skill is stable, offer to refine the description for better triggering.

• Draft 8–10 should-trigger prompts (different phrasings, implicit intents, edge cases).
• Draft 8–10 should-not-trigger prompts (near-misses: same domain but different need).
• Make prompts concrete (paths, context, real-looking detail), not generic.
• Review with the user, then try a few prompts in conversation and see if the skill is suggested when it should be; adjust description wording based on results.

No custom tooling required — this is manual try-and-adjust in Cursor.

Reference

• Cursor skill layout and conventions: see the create-skill skill (~/.cursor/skills-cursor/create-skill/SKILL.md or project equivalent).
• Original full workflow (subagents, eval viewer, benchmark): anthropics/skills – skill-creator.

Summary: Clarify intent → draft SKILL.md (with clear description and structure) → run 2–3 test prompts → collect feedback → improve skill and repeat. Use evals.json and description tuning only when they help.