anthropic-skill-creator（Anthropic 技能 creator）

Create Skills

Build skills that trigger correctly, execute reliably, and stay concise.

Core Principles

• Use progressive disclosure:

- Frontmatter (name, description) for trigger selection. - SKILL.md body for core workflow. - references/ for deep details loaded only when needed.

• Keep skills composable: avoid assumptions that this is the only active skill.
• Prefer deterministic scripts for fragile validation logic.

Workflow

1. Define 2-3 concrete use cases.
2. Draft trigger-strong frontmatter.
3. Design reusable resources (scripts/, references/, assets/).
4. Write actionable instructions and error handling.
5. Run triggering/functional/performance tests.
6. Iterate based on under-trigger/over-trigger signals.

1) Define Use Cases

Capture for each use case:

• User phrasing (what they actually say)
• Intended result
• Multi-step workflow and required tools (built-in and/or MCP)
• Failure points that need guardrails

If use cases are vague, ask targeted follow-up questions before authoring.

2) Frontmatter Design (Most Important)

Write only what is needed to trigger correctly.

Required fields:

• name: kebab-case, matches folder name
• description: include BOTH

- What the skill does - When to use it (trigger phrases/context/file types)

Description formula:

• [What it does] + [When to use it] + [Key capabilities]

Good pattern:

• "Processes PDF legal documents for contract review. Use when users ask for clause extraction, risk flags, or redline summaries from.pdf contracts."

Bad pattern:

• "Helps with projects."

Precision controls:

• Add negative triggers to reduce over-triggering.
• Mention relevant file types when applicable.
• Keep wording aligned with real user phrases from use cases.

3) Resource Plan

• scripts/: deterministic checks/transforms used repeatedly
• references/: large docs, domain logic, variant-specific guidance
• assets/: templates and artifacts used in outputs

Rules:

• Keep SKILL.md compact; link to references/.
• Avoid extra docs in skill folder (README.md, changelog, etc.).

4) Instruction Writing

Write imperative, testable steps.

Include:

• Clear step order and dependencies
• Validation checkpoints
• Common errors with concrete fixes
• Example user inputs and expected outputs

For critical checks, prefer scripts over prose-only checks.

5) Test Protocol

Run these three test groups.

A. Triggering Tests

• Should trigger on obvious requests
• Should trigger on paraphrased requests
• Should NOT trigger on unrelated requests

Target benchmark:

• Trigger on ~90% of relevant prompts in a 10-20 prompt suite

B. Functional Tests

Validate:

• Correct outputs
• Tool/API call success
• Error handling paths
• Edge cases

C. Performance Comparison

Compare with vs without skill:

• Number of clarification turns
• Failed tool/API calls
• Token usage
• End-to-end completion quality

6) Iteration Rules

Under-triggering signals:

• Skill does not load when expected
• Users keep manually invoking it

Fix:

• Add clearer trigger phrases and technical terms to description.

Over-triggering signals:

• Skill loads for unrelated requests

Fix:

• Add negative triggers and tighten scope language.

Execution issues:

• Inconsistent outputs, retries, user corrections

Fix:

• Tighten instructions, add explicit validations, and script critical checks.

Troubleshooting Quick Guide

• Upload fails with missing SKILL.md: file must be exactly SKILL.md.
• Invalid frontmatter: verify YAML delimiters --- and valid syntax.
• Invalid skill name: use lowercase kebab-case.
• Skill not followed: move critical instructions to top and make them concrete.
• Slow behavior: move long content to references/; reduce enabled skills.

Scaffolding

Scaffold a new skill directory:

# Use default path (.agents/skills)
bash scripts/new_skill.sh <skill-name>

# Place in .claude/skills directory
bash scripts/new_skill.sh <skill-name> --path .claude/skills

References

• references/checklist.md
• references/prompt-templates.md
• references/patterns.md
• references/source.md: source PDF and extracted principles
• references/trigger-examples.md: trigger/non-trigger examples for description tuning
• references/failure-remedies.md: failure patterns and minimum effective fixes