claude-extensibility（Claude extensibility 搜索）

Claude Code Extensibility

CRUD operations for agents, skills, and output styles. This skill transfers expertise in building extensions that actually activate and deliver value.

Related Skills

IMPORTANT: When creating or editing prompts, use prompt-architect skill to improve quality.

Skill("prompt-architect")  → Create/enhance skill/agent prompt content

Core Principles

• Simplicity: Direct tool calls, avoid complex abstractions
• Focus: Single, clear responsibility per extension
• Conciseness: Target <800 lines, use progressive disclosure
• Efficiency: Optimize for token usage and response time

Core Truths

Every extension must embody these principles:

Decision Heuristics

Expert Patterns

The "Everything Tool" Request

When you see: User wants one agent/skill to handle auth, database, API, and testing This indicates: Scope creep that will create bloated, unfocused extension Therefore: Split into focused single-responsibility extensions; suggest composition Watch out: User may resist - explain that 3 focused skills outperform 1 swiss-army-knife

The Wrapper Trap

When you see: Skill that just wraps a CLI tool with no added expertise This indicates: Missing the expertise transfer opportunity Therefore: Ask "what would an expert know that the tool doesn't tell you?" Watch out: The skill becomes redundant with just reading --help

Activation Keyword Starvation

When you see: Frontmatter description uses only formal/technical terms This indicates: Skill won't activate when users use casual language Therefore: Add synonyms, verbs, casual phrasings to description Watch out: Over-broad keywords cause false activations

The Procedure Manual

When you see: Step 1, Step 2, Step 3... (long numbered instructions) This indicates: Procedures don't transfer judgment - expert skips steps, adapts order Therefore: Teach patterns and heuristics. Let reader apply judgment. Watch out: Some workflows genuinely need sequence - use sparingly

Extension Types

Activation Keyword Engineering

The description field is your activation gate. System reads ONLY frontmatter to decide loading. Brilliant skill + poor keywords = dead skill.

Keyword Categories to Include

1. Task Verbs: create, build, debug, fix, deploy, optimize, refactor
2. Problem Descriptions: slow, broken, failing, error, conflict
3. Artifact Types: component, API, database, test, config
4. Tool/Framework Names: React, PostgreSQL, Docker (exact + misspellings)
5. Casual Synonyms: "make it faster" → optimize, performance

Description Format (WHAT + WHEN)

# Bad - too vague, won't trigger
description: Helps with documents

# Bad - missing WHEN triggers
description: PDF skill

# Good - multiple trigger points
description: "Extract text and tables from PDF files, fill forms, merge documents.
  Formats: .pdf. Tools: pypdf, pdfplumber. Actions: extract, fill, merge PDFs.
  Keywords: PDF, form, document. Use when: working with PDF files, extracting data."

Frontmatter Checklist

• Name is clear and descriptive (not abbreviation)
• Description includes file types if applicable (.docx,.pdf)
• Description includes task verbs (creating, editing, analyzing)
• Description includes synonyms users might say
• Description includes contexts (reports, professional documents)

Keyword Density Test

Read description aloud. For every 10 words, at least 3 should be potential triggers.

Agent Development

Reference: references/agent-development.md - Full YAML structure, system prompt patterns, optimization techniques.

Quick Start: Agent

---
name: agent-name
# CRITICAL: Description is the ONLY thing system reads to decide activation
# Include: task verbs, problem descriptions, synonyms users actually say
description: Use this agent when [use case]. Use PROACTIVELY for [triggers].\n\nExamples:\n<example>\nContext: [situation]\nuser: [request]\nassistant: [response]\n<commentary>[reasoning]</commentary>\n</example>
# CONSTRAIN RUTHLESSLY: 3-5 tools = focused. 10+ tools = confused.
tools: Grep, Glob, Read, Bash
model: haiku
permissionMode: default
skills: skill-name
---

# Agent Name

Brief mission statement.

## Core Strategy

### 1. Phase Name
Approach and techniques

## What This Agent Does NOT Do
- Does not [boundary 1] (use X agent)
- Does not [boundary 2] (escalate to user)

<format>
Expected output structure
</format>

YAML Fields

Model Selection

Built-in Subagents

Agent Locations

CLI-Defined Agents

claude --agents '{
  "code-reviewer": {
    "description": "Expert code reviewer. Use proactively after code changes.",
    "prompt": "You are a senior code reviewer...",
    "tools": ["Read", "Grep", "Glob", "Bash"],
    "model": "sonnet"
  }
}'

Resumable Agents

Continue previous conversations:

• Each execution gets unique agentId
• Transcript stored in agent-{agentId}.jsonl
• Resume with previous agentId to continue with full context

Skill Development

Reference: references/skill-development.md - Full structure, trigger patterns, hook system.

Quick Start: Skill

---
name: skill-name
# WHAT it does + WHEN to use it (both required!)
description: "[Core purpose]. [Technologies]. Capabilities: [list].
  Actions: [verbs]. Keywords: [triggers]. Use when: [scenarios]."
allowed-tools: Read, Grep, Glob
---

# Skill Name

## Purpose
What this skill helps with - in practitioner voice.

## Patterns

### [Pattern Name]
**When you see:** [Observable trigger]
**This indicates:** [Expert insight]
**Therefore:** [Action to take]
**Watch out for:** [Common pitfall]

## Heuristics
- **[Rule of thumb]**: [When/why it applies]

## When to Use Something Else
- For X → use `other-skill`
- For Y → escalate to user

YAML Fields

Tool Access Control

Restrict Claude's tools with allowed-tools:

---
name: safe-reader
description: "Read-only file access. Use when viewing code without modifications."
allowed-tools: Read, Grep, Glob
---

Skill Locations

Skill Structure

my-skill/
├── SKILL.md (required, <800 lines)
├── references/ (optional - detailed docs)
├── scripts/ (optional - utilities)
└── templates/ (optional - templates)

Core Sections (Include What's Needed)

Output Styles

Modify Claude Code's main agent behavior.

Quick Start: Output Style

---
name: My Custom Style
description: Brief description of behavior
keep-coding-instructions: true
---

# Custom Style Instructions

You are an interactive CLI tool that helps users...

## Specific Behaviors
[Define assistant behavior...]

YAML Fields

Built-in Styles

• Default: Standard software engineering
• Explanatory: Educational insights between tasks
• Learning: Collaborative with TODO(human) markers

Output Style Locations

• User: ~/.claude/output-styles/
• Project: .claude/output-styles/

Usage

/output-style              # Access menu
/output-style explanatory  # Switch directly

Writing Principles

Voice

• Sound like a practitioner, not documentation
• Direct and confident, not hedging
• Specific and concrete, not abstract

Before (documentation voice):

It is recommended that users consider implementing appropriate error handling mechanisms.

After (practitioner voice):

Always handle errors explicitly. Silent failures are debugging nightmares.

Density

• Every sentence should transfer knowledge
• No filler, no redundancy
• If it doesn't change behavior, cut it

Scannability

• Claude reads this during tasks
• Headers help Claude find relevant sections
• Patterns are scannable by name
• Bold key terms

Patterns Over Procedures

Transform procedures to patterns when expertise matters:

Before (procedure):

1. Open the file
2. Check the format
3. Validate the data
4. Process the content

After (pattern):

### Format Recognition
**When you see:** File extension and initial bytes
**This indicates:** Expected structure and parsing approach
**Therefore:** Choose parser before reading content
**Watch out for:** Extension doesn't always match actual format

Anti-Patterns

The Kitchen Sink Agent

Looks like: Agent with 10+ tools, handles "everything related to X" Why wrong: Broad agents make poor decisions, consume tokens exploring options Instead: Constrain to 3-5 core tools. Create sibling agents for related domains.

The Echo Skill

Looks like: Skill that restates tool documentation without adding insight Why wrong: No expertise transfer. User could just read the docs. Instead: Add the "expert layer" - pitfalls, non-obvious combos, when NOT to use.

The Invisible Trigger

Looks like: Description uses only official tool/framework name Why wrong: Users say "help me deploy" not "invoke kubernetes-orchestrator" Instead: Include task verbs, user-language synonyms, problem descriptions.

Missing Boundaries

Looks like: No explicit "what this does NOT do" section Why wrong: Scope creep, incorrect expectations, overlapping extensions Instead: Every extension needs 3+ explicit out-of-scope declarations.

Vague Descriptions

Looks like: description: Helps with documents Why wrong: Too vague to trigger on relevant requests Instead: Include file types, task verbs, problem descriptions, synonyms.

Quality Signals

Good Signs

• Frontmatter reads like "when to use" guide, not feature list
• First 10 lines provide immediate actionable guidance
• Expert reading it would nod and say "yes, that's how I think"
• Explicit constraints on what it does NOT handle
• Examples show judgment calls, not just syntax
• Has named patterns with insights, not just steps

Warning Signs

• More than 50% content is reference tables or field definitions
• No mention of tradeoffs or "when NOT to use"
• Generic language that could apply to any domain
• Activation keywords are all nouns, no verbs or problem descriptions
• Body exceeds 600 lines (probably doing too much)
• Second-person voice ("you should...")

Validation Checklist

Before outputting, verify against Core Truths:

Frontmatter (Most Critical)

• Has YAML frontmatter with name and description?
• Name is clear and searchable (not abbreviation)?
• Description includes trigger keywords?
• Description includes file types if applicable?
• Description includes task verbs (creating, editing, etc.)?
• Description includes synonyms users might use?

Expertise Transfer

• Has named patterns with insights, not just steps?
• Transfers HOW to think, not just WHAT to do?
• Would an expert recognize their thinking here?

Flow

• Helps produce output, not create busywork?
• No unnecessary intermediate steps?
• Actionable immediately?

Focus

• Every section earns its place?
• No bloat?
• Ruthlessly constrained to what matters?

Common Workflows

Create Agent

1. Create .claude/agents/{name}.md
2. Write YAML frontmatter (name, description, tools, model)
3. Write system prompt (<800 lines)
4. Use Skill("prompt-architect") to improve prompt
5. Test with Task tool
6. Optimize based on performance

Create Skill

1. Create .claude/skills/{name}/SKILL.md
2. Write YAML frontmatter with WHAT + WHEN description
3. Write content with patterns, not procedures
4. Use Skill("prompt-architect") to improve prompt
5. Add reference files for detailed content
6. Test: Does it activate when expected?

Optimize Extension

1. Measure baseline (lines, token usage, response time)
2. Move details to reference files
3. Use Skill("prompt-architect") to improve prompts
4. Remove second-person voice
5. Use code blocks over prose
6. Add XML structure
7. Test and verify improvements

Testing

Key Question: Does it activate when expected?

Agent Testing:

Task(
  subagent_type="agent-name",
  description="Test task",
  prompt="Detailed test prompt"
)

Skill Testing:

• Test prompts that SHOULD trigger
• Test prompts that should NOT trigger
• Debug with: claude --debug

Best Practices

Anthropic Guidelines

✅ 800-line rule: Keep SKILL.md and agent prompts under 800 lines ✅ Progressive disclosure: Use reference files for detailed content ✅ Proactive language: Include "use PROACTIVELY" in descriptions ✅ WHAT + WHEN descriptions: Both capability and triggers ✅ Test first: Build 3+ evaluations before extensive documentation ✅ Least privilege: Limit tools to necessary set

Anti-Pattern Summary

❌ Vague descriptions without triggers ❌ Over 800 lines without references ❌ Second-person voice ("you should...") ❌ All tools when subset suffices ❌ No examples in agent descriptions ❌ Procedures without patterns

Quick Reference

Agent Model Selection:

• Haiku: Fast, simple tasks (< 3s)
• Sonnet: Balanced, most use cases (< 10s)
• Opus: Complex reasoning (< 30s)
• Inherit: Match main conversation

File Locations:

• Agents: .claude/agents/*.md
• Skills: .claude/skills/{name}/SKILL.md
• Output Styles: .claude/output-styles/*.md

Management Commands:

• /agents - Interactive agent management
• /output-style - Switch output styles

Status: Production Ready | Lines: ~500 | Progressive Disclosure: ✅