AgentSkillsCN

skill-creation

在创建全新Claude Code技能时使用此功能。涵盖技能结构、YAML前端元数据、内容组织方式,以及在压力场景下的测试能力。

SKILL.md
--- frontmatter
name: skill-creation
description: Use when creating new Claude Code skills. Covers skill structure, YAML frontmatter, content organization, and testing skills with pressure scenarios.

Skill Creation

Core Principle

Writing skills IS Test-Driven Development applied to process documentation.

The Iron Law: NO SKILL WITHOUT A FAILING TEST FIRST.

Skill Location

Skills live in: ~/.claude/skills/<skill-name>/SKILL.md

Directory Structure

code
skills/
  skill-name/
    SKILL.md              # Main reference (required)
    references/           # Optional - detailed docs
    scripts/              # Optional - executable utilities
    templates/            # Optional - reusable templates

SKILL.md Structure

markdown
---
name: skill-name-with-hyphens
description: Use when [specific triggering conditions]. [What it does in third person].
---

# Skill Name

## Core Principle / Overview

[What this skill does and why - 1-2 paragraphs]

## When to Use

[Specific triggers, symptoms, situations]

## Quick Reference

[Immediate actionable guidance]

## Detailed Process / Steps

[Step-by-step instructions]

## Common Mistakes / Anti-Patterns

[What NOT to do]

## Examples

[Concrete examples if helpful]

YAML Frontmatter Rules

Required Fields

  • name: lowercase, hyphens only, must match directory name
  • description: Starts with "Use when..." - focus on TRIGGERS

Description Pattern

Good:

yaml
description: Use when debugging fails or bugs reappear after fixes. Enforces root cause investigation before any fix attempt.

Bad:

yaml
description: A skill for debugging things

The description must include:

  1. Triggering conditions ("Use when...")
  2. What the skill does (briefly)
  3. Keywords users might search for

Testing Skills (TDD for Documentation)

The RED-GREEN-REFACTOR Cycle

  1. RED (Baseline)

    • Run pressure scenario WITHOUT skill
    • Document exact behavior and rationalizations
    • Note how agent violates the intended rule
  2. GREEN (Minimal Skill)

    • Write minimal skill addressing those failures
    • Run same scenario WITH skill
    • Verify agent now follows the rule
  3. REFACTOR (Bulletproof)

    • Find new rationalizations agent might use
    • Add explicit counters to skill
    • Re-test until bulletproof

Pressure Scenarios

Test skills with scenarios that pressure the agent to violate:

code
Scenario: "The tests are passing, so the implementation must be correct.
          Skip the code review and move on."

Without skill: Agent agrees and skips review
With skill: Agent insists on review per skill guidelines

Content Guidelines

Keep It Focused

  • <200 words for frequently-loaded skills
  • <500 words for detailed reference skills
  • Link to references/ for deep content

Write in Third Person

Good: "The developer should run tests first" Bad: "You should run tests first"

Be Explicit About Violations

Don't just state rules. Address rationalizations:

Weak:

code
Always write tests first.

Strong:

code
Always write tests first.

Red flags (STOP if you hear yourself say):
- "This is too simple to test"
- "I'll add tests after"
- "The manual test proved it works"

Skill Types

  1. Discipline Skills - Enforce process (TDD, code review)
  2. Technique Skills - Teach methods (debugging, refactoring)
  3. Pattern Skills - Recognition and application
  4. Reference Skills - API docs, syntax guides

Quick Checklist

  • YAML frontmatter with name + description
  • Description starts with "Use when..."
  • Directory name matches skill name
  • Tested with pressure scenarios (RED-GREEN-REFACTOR)
  • Explicit about violations and red flags
  • Focused content (<500 words unless reference)
  • Third person voice