skill-creator（技能创建器）

Creating Skills and Commands

This skill provides guidance for creating effective skills and commands in Claude Code projects.

Skills vs Commands

Core Principles

Concise is Key

The context window is a public good. Skills share it with system prompts, conversation history, and the user's actual request.

Default assumption: Claude is already very smart. Only add context Claude doesn't already have. Challenge each piece: "Does Claude really need this?"

Prefer concise examples over verbose explanations.

Set Appropriate Degrees of Freedom

Match specificity to the task's fragility:

High freedom (text instructions): Multiple approaches valid, context-dependent

Medium freedom (pseudocode/templates): Preferred pattern exists, some variation OK

Low freedom (specific scripts): Operations fragile, consistency critical

Skill Structure

skill-name/
├── SKILL.md (required)
│   ├── YAML frontmatter (name, description - required)
│   └── Markdown instructions
└── Optional resources
    ├── references/    # Documentation loaded as needed
    ├── scripts/       # Executable code
    └── assets/        # Templates, images, etc.

SKILL.md Frontmatter

---
name: skill-name
description:
  What this skill does and when to use it. This is the PRIMARY trigger -
  Claude reads this to decide when to load the skill. Include specific
  scenarios and keywords that should activate it.
---

Important: The description is loaded into context ALWAYS. The body is only loaded AFTER the skill triggers. Put "when to use" in the description, not the body.

Command Structure

Commands are simpler - single markdown files:

.claude/commands/
├── daily-review.md
├── release.md
└── new-command.md

Command Frontmatter

---
name: command-name
description: 'Brief description shown in /help'
argument-hint: '[optional] [args]'
allowed-tools: [Read, Glob, Bash(git:*)]  # Optional: restrict tools
---

Progressive Disclosure

Use layered loading to manage context efficiently:

1. Metadata (name + description) - Always in context (~100 words)
2. SKILL.md body - When skill triggers (<500 lines ideal)
3. References - As needed by Claude (unlimited)

Pattern: Reference Files

Keep SKILL.md lean, link to detailed references:

# Skill Name

## Quick Reference

Basic usage here.

## Detailed Guides

- **Complex Topic A**: See `references/topic-a.md`
- **Complex Topic B**: See `references/topic-b.md`

Claude reads reference files only when needed.

Creating a New Skill

Step 1: Understand the Use Case

Ask:

• "What functionality should this skill support?"
• "What would trigger someone to need this?"
• "What does Claude NOT already know that this should teach?"

Step 2: Plan the Contents

Analyze concrete examples:

• What code/scripts get rewritten each time?
• What documentation would help?
• What assets (templates, examples) are needed?

Step 3: Create the Skill

mkdir -p .claude/skills/my-skill

Write SKILL.md:

---
name: my-skill
description:
  What it does. When to use it. Keywords that trigger it.
---

# My Skill

## Overview

Brief explanation.

## Quick Reference

Essential information here.

## Advanced Topics

Link to references if complex.

Step 4: Test and Iterate

1. Start a new Claude session
2. Ask questions that should trigger the skill
3. Verify Claude uses the skill appropriately
4. Refine based on gaps

Creating a New Command

Step 1: Define the Workflow

Commands are for explicit multi-step workflows:

• What steps does this workflow include?
• What tools are needed?
• What output should the user see?

Step 2: Create the Command

touch .claude/commands/my-command.md

Write the command:

---
name: my-command
description: 'Brief description for /help'
argument-hint: '[--flag] [arg]'
---

# My Command

## Instructions

1. First step
2. Second step
3. Verify result

## Handling Edge Cases

What to do when X happens.

Step 3: Test

Run /my-command and verify the workflow executes correctly.

Common Patterns

Workflow with Verification

## Workflow

1. Gather context
2. Perform action
3. **Verify result before claiming success**
4. Report outcome

## Verification

Always run `<command>` and check output before claiming complete.

Reference Hub

# API Reference Hub

| API | Purpose | Reference |
|-----|---------|-----------|
| API A | Feature X | `references/api-a.md` |
| API B | Feature Y | `references/api-b.md` |

## When to Use Each

Brief guidance on which reference to consult.

Decision Matrix

## Approach Selection

| Condition | Approach |
|-----------|----------|
| Simple case | Do X |
| Complex case | Do Y |
| Edge case | Ask user |

What NOT to Include

• README.md, CHANGELOG.md (just clutter)
• Obvious information Claude already knows
• Overly detailed explanations (use examples instead)
• User-facing documentation (skills are for Claude)

Skill vs Command Decision

Use a Skill when:

• Providing background knowledge
• Teaching domain expertise
• Reference material that applies across tasks

Use a Command when:

• Interactive workflow with user
• Multi-step process needing explicit invocation
• Specific task with defined start/end