creating-skills（创造技能）

Creating Skills

This skill walks you through creating a well-structured agent skill from scratch. Follow the workflow below step by step. Use the reference files for detailed guidance on specific topics.

Reference Files

Core Principles

Before diving into mechanics, internalize these design principles:

Context window is a public good. Every token in a skill competes with the user's code, conversation, and other tools. Challenge each line with three questions: (1) Does Claude already know this? (2) Will the agent use this on every invocation? (3) Can this live in a reference file instead?

Degrees of freedom. Match your specificity to the task's tolerance for variation:

Claude is already smart. Don't teach programming, well-known APIs, or common patterns. Only add context Claude doesn't already have: project conventions, domain rules, non-obvious constraints.

Enforce mechanically, not with checklists. If the skill's purpose is catching problems (review, audit, compliance), it should generate hooks, lint rules, or check scripts that run automatically — not be a list of things the agent must remember to check. A PostToolUse hook runs on every file edit; a skill instruction relies on agent discipline.

Don't duplicate existing tools. Before building a skill, check what linters, formatters, and CI tools already handle in the target ecosystem. If ESLint catches unused imports, don't create a skill that also checks for them. Skills should add value beyond what standard tooling provides.

Design for discoverability. A skill that never gets invoked is worthless. The description's "Use when" triggers are necessary but not sufficient — the agent scans them at session start and may not connect them to the right moment. Add cross-references in related skills ("Related Skills" section) and in CLAUDE.md for broadly useful skills. Ask: "When would someone need this, and what will they be looking at right before they need it?"

Choose a Skill Pattern

Before writing anything, identify which pattern fits the task:

Not sure? Ask these questions:

1. Does the task have sequential steps? → Guided Workflow
2. Is it about checking/validating? → Rules-Based Audit
3. Does it produce files from templates? → Scaffolding / Generation
4. Is it primarily informational? → Knowledge Reference

Read references/skill-patterns.md for full skeleton templates and directory structures.

Creation Workflow

Copy this checklist and work through it step by step:

- [ ] Step 1: Understand the skill
- [ ] Step 2: Choose a pattern
- [ ] Step 3: Initialize the skill
- [ ] Step 4: Write the SKILL.md body
- [ ] Step 5: Add reference files
- [ ] Step 6: Validate
- [ ] Step 7: Create evaluations
- [ ] Step 8: Test with real usage
- [ ] Step 9: Iterate

Step 1: Understand the Skill

Before writing any files, have a conversation to clarify:

• What does this skill do? Get a one-sentence answer.
• Who triggers it? What prompts or situations should activate it?
• What's the output? Files created, code reviewed, information provided?
• What does the agent need to know? Domain knowledge, project conventions, constraints?
• What does the agent already know? Don't include common programming knowledge.

Synthesize into a draft description using the formula: [Does what] for/using [domain]. [Checks/covers what]. Use when [triggers]

Step 2: Choose a Pattern

Use the decision table above to select a primary pattern. Most skills combine patterns — pick the dominant one and incorporate elements from others.

For example, creating-skills is primarily a Guided Workflow but includes:

• Rules-Based Audit elements (validation checklist)
• Scaffolding elements (init script)

Step 3: Initialize the Skill

Quick start with the official CLI:

npx skills init <skill-name>

Or use the enhanced scaffolding script (includes best-practice guidance in the template):

python scripts/init_skill.py <skill-name> --path skills

This creates:

skills/<skill-name>/
├── SKILL.md          (template with TODOs)
└── references/       (empty, ready for use)

If not using the script, create this structure manually. The name must be kebab-case, preferably with a gerund first word (e.g., analyzing-data, not data-analyzer).

Step 4: Write the SKILL.md Body

Open the generated SKILL.md and fill it in:

1. Frontmatter: Fill in name and description

- Name: kebab-case, must match directory name - Description: third-person voice, recommend under 300 characters (spec max 1024), include "Use when" triggers - See references/format-specification.md for the full spec

1. Reference Files table: List every reference file with a "Read When" condition
2. Main content: Follow the skeleton for your chosen pattern

- Target 150–300 lines for the body - Hard limit: 500 lines (agents lose focus beyond this) - Use progressive disclosure — put detailed content in reference files - Use ATX headings, fenced code blocks with language tags - One sentence per line for clean diffs

1. Workflows: Use copyable checklists for sequential steps

- See references/workflow-and-output-patterns.md for patterns

Step 5: Add Reference Files

Create reference files in references/ for detailed content that would bloat the main SKILL.md.

Common reference file types:

• Rules and specifications — naming conventions, format constraints, allowed values
• Examples — input/output pairs, good/bad comparisons
• Detailed step guides — expanded instructions for complex steps
• Decision matrices — comparison tables for choosing between options

Rules for reference files:

• One level deep only (no nested directories under references/)
• Every file must be listed in the SKILL.md "Reference Files" table
• Files over 100 lines should include a table of contents
• Use kebab-case file names

Step 6: Validate

Run the validator against your skill:

python scripts/validate_skill.py skills/<skill-name>

The validator checks:

• Frontmatter format and required fields
• Name matches directory, kebab-case format
• Description length and voice
• Body line count
• File references exist on disk
• No Windows-style paths
• No deeply nested references
• Reference files are listed in the body

For official spec compliance, also run: skills-ref validate skills/<skill-name> (install: pip install skills-ref)

Fix all errors (blocking). Review all warnings (advisory — fix or acknowledge).

If errors persist after 3 attempts, review references/anti-patterns.md for common mistakes.

Also run through the manual references/quality-checklist.md for items the automated validator cannot check (content quality, token efficiency, terminology consistency).

Step 7: Create Evaluations

Before testing, define what success looks like. Write 3–5 evaluation cases that cover the happy path, an edge case, and a failure mode you anticipate. See references/evaluation-guide.md for the full methodology and JSON format.

Key steps:

1. Establish a baseline: Complete the target task without the skill installed. Record where the agent struggles.
2. Write evaluation cases: Each is a prompt + expected behaviors checklist.
3. Score the baseline: This is the bar your skill must beat.

Step 8: Test with Real Usage

Automated validation catches format issues but not usability problems. Test with real tasks:

1. Install the skill: Copy to ~/.claude/skills/ or use npx skills add
2. Start a fresh session: The agent should not have prior context about this skill
3. Trigger naturally: Use a prompt that would naturally activate the skill
4. Evaluate the output: Did the agent follow the workflow? Was the output correct?

What to observe during testing:

• Does the agent explore unexpected paths not covered by the skill?
• Does the agent miss connections between sections that seem obvious to you?
• Does the agent over-rely on one section while ignoring others?
• Is there content the agent consistently skips or ignores?

Multi-model testing: Test with at least two capability levels:

• Haiku: Can the skill work with a smaller model? Simplify if not.
• Sonnet/Opus: Does the skill produce high-quality output with a capable model?

If the skill fails or produces poor output, go back to Step 4 and revise.

Step 9: Iterate

Skills improve through real usage and feedback.

Establish a baseline first: If you haven't already (Step 7), complete the target task without the skill to see what the agent does on its own. This reveals which parts of your skill actually add value vs. what the agent already handles.

Use the Claude A/B pattern:

1. Claude A: Use the skill as-is with a real task. Note where it struggles.
2. Claude B: Open a new session. Describe the problems. Ask Claude B to suggest improvements to the SKILL.md.
3. Apply and re-test: Incorporate the suggestions, then repeat from Step 6.

Gather team feedback: If others use the skill, ask them what worked and what didn't. Different users trigger skills differently — their experience reveals gaps your testing missed.

This external feedback loop catches blind spots that self-review misses.

Anti-Patterns Quick Reference

Avoid these common mistakes (see references/anti-patterns.md for full details):

Related Skills

When using this skill alongside others:

• After creating a skill, use a reviewing skill to audit it against project standards
• For skills that generate code, pair with a testing skill to validate output
• For API-related skills, consider a documenting skill for endpoint coverage