skill-creator（技能创建器）

Skill Authoring

Create agent skills (SKILL.md files) following the Agent Skills open standard. Skills produce consistent, deterministic agent behavior across any compatible platform. Skills are filesystem-based instruction sets that transform an agent into a domain specialist.

When invoked with arguments, create a skill for: $ARGUMENTS

Workflow

1. Define scope: Identify the domain, target tasks, and activation triggers
2. Gather requirements: Collect domain knowledge, constraints, expected inputs/outputs, and success criteria. Ask clarifying questions for missing context
3. Determine freedom levels: Classify each instruction as high/medium/low freedom based on task sensitivity
4. Write frontmatter: Name, description with activation triggers
5. Write instructions: Imperative commands, structured for LLM parsing
6. Add examples: Concrete input/output pairs for ambiguous or complex behaviors
7. Add supporting files: Move detailed reference material to separate files when SKILL.md approaches 300 lines
8. Verify: Run through the checklist below

SKILL.md Structure

Every skill requires a SKILL.md file in a named directory:

skill-name/
├── SKILL.md              # Required. Main instructions (under 500 lines)
├── references/           # Optional. Deep-dive material
│   └── *.md              # One level deep only
├── scripts/              # Optional. Executable utilities
│   └── *.py|*.sh
└── assets/               # Optional. Templates, schemas, static resources

Frontmatter

YAML frontmatter between --- markers. The open standard requires name and description.

---
name: kebab-case-name
description: What this skill does. Use when (1) trigger one, (2) trigger two, (3) trigger three, (4) trigger four, or (5) trigger five.
---

Name (required)

Must match the parent directory name. Lowercase letters, numbers, hyphens only. Maximum 64 characters. Must not start or end with a hyphen. Must not contain consecutive hyphens (--).

Description (required)

Maximum 1024 characters. Write in third person. Include both what the skill does AND when to activate.

Template: [Value proposition sentence]. Use when (1) [trigger], (2) [trigger],... (N) [trigger].

Enumerate 4-6 specific activation triggers. Base triggers on user intent and tasks, not file types.

Optional Universal Fields

For platform-specific extensions (Claude Code's argument-hint, context, agent, disable-model-invocation, user-invocable, model, hooks), see references/frontmatter.md.

Writing Instructions

Core Principles

The context window is a shared resource. Only add context the model lacks. Assume high baseline capability.

Imperative language only. "Validate input at boundaries" not "You should consider validating input."

Zero ambiguity. Every instruction must have exactly one interpretation. Eliminate hedge words: "consider", "try to", "when possible", "generally".

Specificity scales with risk. Use high freedom (text instructions) when multiple approaches are valid. Use low freedom (exact commands, scripts) when consistency is critical.

Formatting for LLM Parsing

Emphasis Modifiers

Use MUST, MUST NOT, REQUIRED only for hard constraints where violation causes failure. If every instruction uses MUST, none stand out. Reserve bold for the same purpose.

Terminology Consistency

Choose one term per concept. Use it throughout the entire skill. Do not alternate between synonyms.

Progressive Disclosure

Keep SKILL.md body under 500 lines. Start with core instructions. Link to separate files for deep-dive material.

Pattern: High-level guide in SKILL.md with [reference.md](reference.md) links for specialized content.

Keep all references one level deep from SKILL.md. Do not chain: SKILL.md -> advanced.md -> details.md.

Reference files longer than 100 lines include a table of contents at the top.

Instruction Patterns

Workflow Pattern

Use for procedural, multi-step tasks. Numbered steps with clear sequencing.

## Workflow
1. Detect framework from `package.json`
2. Select base image from Image Guide table
3. Generate Dockerfile using selected pattern
4. Create `.dockerignore` if missing
5. Validate with checklist

Decision Table Pattern

Use for selection logic with discrete options.

| Condition | Action |
|-----------|--------|
| API endpoint | Validate input schema, return JSON |
| Background job | Log start/end, handle timeout |
| CLI command | Parse args with yargs, exit codes 0/1 |

Template Pattern

Use when output must follow an exact structure.

Output format:

[Title]

Summary

[2-3 sentence overview]

Findings

Conditional Workflow Pattern

Use when the workflow branches based on input.

1. Determine modification type:
   **Creating new?** -> Follow Creation workflow
   **Editing existing?** -> Follow Editing workflow

Examples Pattern

Use input/output pairs to demonstrate expected behavior. Wrap in <example> tags for complex cases.

<example>
Input: User requests API endpoint for user registration
Output:

// POST /api/users export async function handler(req: Request) { const body = await req.json() // validate, create user, return 201 }

</example>

For additional patterns (feedback loops, checklist tracking, dynamic context), see references/patterns.md.

Content Guidelines

What to Include

• Domain-specific knowledge the model lacks
• Exact commands with all flags and arguments
• Concrete examples demonstrating desired output format
• Decision tables for selection logic
• Verification steps for critical operations
• Constraints and prohibited actions

What to Exclude

• Information the model already knows (framework basics, language syntax)
• Decorative language ("please", "remember", "it's important")
• Welcome messages, motivational text, background history
• Time-sensitive information ("Before August 2025, use...")
• Hypothetical scenarios ("If we ever migrate to...")
• Vague best practices ("Functions should be small")

Reducing Hallucination Risk

When a skill processes external documents or data:

• Instruct to extract direct quotes before analyzing
• Require citations for claims
• Include explicit fallback: "If information is unavailable, state 'Insufficient data' instead of inferring"
• Restrict to provided context when accuracy is critical

For comprehensive anti-patterns reference, see references/anti-patterns.md.

Platform Extensions

The open standard defines the universal format above. Individual platforms add extensions for deeper integration. Use universal fields by default; add platform extensions only when the extra capability is needed.

Claude Code Extensions

Skills double as slash commands in Claude Code. Support argument passing with $ARGUMENTS.

---
name: review-pr
description: Review a pull request for code quality and standards
argument-hint: "[PR number or URL]"
---
Review pull request $ARGUMENTS.
1. Fetch PR diff
2. Check against coding standards
3. Report findings

Positional arguments: $0, $1, $2 or $ARGUMENTS[0], $ARGUMENTS[1].

Dynamic context injection runs shell commands and inserts output before the model sees the content. Prefix a backtick-wrapped command with the bang character (the exclamation mark).

For complete syntax reference and a working example, see references/frontmatter.md.

Other platforms may define their own extensions. Consult the target platform's documentation for available fields.

For the complete Claude Code extension reference, see references/frontmatter.md.

Verification Checklist

Run through before finalizing any skill.

Frontmatter (Universal)

• name is kebab-case, under 64 characters, matches parent directory name
• name does not start/end with hyphen, no consecutive hyphens
• description is non-empty, under 1024 characters, third person
• Description includes what the skill does AND 4-6 activation triggers
• license field present if skill will be shared or published
• compatibility field present if skill requires specific environment

Frontmatter (Platform-Specific)

• argument-hint present if skill accepts slash command arguments (Claude Code)
• Platform-specific fields only used when universal fields are insufficient

Instructions

• All instructions use imperative mood
• No hedge words (consider, try to, when possible, generally)
• No decorative language (please, remember, make sure)
• Zero ambiguity: each instruction has exactly one interpretation
• Emphasis modifiers (MUST, bold) reserved for hard constraints only
• Consistent terminology: one term per concept throughout
• Code blocks are language-labeled

Structure

• SKILL.md body under 500 lines
• Sections follow: Title -> Overview -> Workflow/Principles -> Detail -> Examples -> Checklist
• Tables used for structured data, lists for discrete items, prose for relationships
• Supporting files one level deep from SKILL.md
• Reference files over 100 lines have a table of contents

Content

• Only adds context the model lacks
• No framework documentation the model already knows
• No time-sensitive information
• Examples are concrete input/output pairs, not abstract descriptions
• Output format explicitly defined when consistency matters
• Verification/validation steps included for critical operations

Execution

• Produces consistent behavior across multiple runs
• Includes clear invocation conditions (when to use AND when NOT to use)
• Specifies expected inputs, outputs, and validation criteria
• Freedom levels match task sensitivity (high freedom for creative tasks, low for critical operations)