Claude Code Knowledge Base

Quick Reference for Formats

Commands

Path:

.claude/commands/name.md

Invocation:

/name

or

/name arguments

---
description: Required. What the command does.
allowed-tools: Optional. Restrict tools.
model: Optional. opus/sonnet/haiku
argument-hint: Optional. Hint for arguments.
---

Command instructions.

Use $ARGUMENTS to insert user arguments.

Examples of good commands:

---
description: Creates a git commit with meaningful message based on staged changes
---

1. Run `git diff --staged`
2. Analyze changes
3. Generate commit message:
   - Title up to 50 characters
   - Empty line
   - Detailed description
4. Run `git commit -m "..."`

---
description: Runs code review for specified file or directory
argument-hint: [path to file or directory]
allowed-tools: Read, Grep, Glob
---

Perform code review for: $ARGUMENTS

Check:
- Code quality
- Potential bugs
- Security
- Performance

Output format:
## Critical Issues
## Warnings
## Recommendations

---

Agents

Path:

.claude/agents/name.md

Invocation: Automatically or "Use agent name for..."

---
name: agent-name  # required
description: Required. When to use the agent.
tools: Optional. All by default.
model: Optional. opus (default) / haiku / sonnet / inherit
permissionMode: Optional. default / acceptEdits / bypassPermissions / plan
skills: Optional. Auto-load skills.
---

Agent system prompt.

Available tools:

• Read, Write, Edit — file operations
• Bash — execute commands
• Grep, Glob — search
• WebSearch, WebFetch — web
• Task — create subagents (not recursive)
• MCP tools — if configured

Examples of good agents:

---
name: researcher
description: Researches codebase and gathers information. Use PROACTIVELY before implementing new features.
tools: Read, Grep, Glob, Bash
model: haiku
---

You are a codebase researcher.

## Task
Quickly find and analyze relevant code.

## Process
1. Glob — find files by pattern
2. Grep — find usages/definitions
3. Read — study key files
4. Summarize findings

## Output
- Found files and their roles
- Code patterns
- Recommendations

---
name: test-writer
description: Creates tests for code. MUST BE USED after writing new functionality.
tools: Read, Write, Bash
model: opus
---

You are a testing specialist.

## Process
1. Read code that needs testing
2. Determine project's test framework
3. Write tests:
   - Unit tests for functions
   - Edge cases
   - Error handling
4. Run tests
5. Fix if failing

---

Skills

Path:

.claude/skills/name/SKILL.md

Invocation:

/name

or automatically

---
name: skill-name  # lowercase, hyphens, max 64
description: Required. What and when. Max 1024 chars.
allowed-tools: Optional. Restrict.
disable-model-invocation: true  # only user invokes
user-invocable: false  # only Claude invokes
---

Skill instructions.

Skill folder structure:

skill-name/
├── SKILL.md        # required
├── scripts/        # executable code
├── references/     # additional documentation
└── assets/         # templates, resources

Example skill with resources:

---
name: api-design
description: REST API design patterns. Use when creating or reviewing API endpoints.
---

# API Design Patterns

## Principles
- RESTful naming
- Consistent error format
- Proper status codes

For detailed examples see [references/examples.md](references/examples.md)
For templates see [assets/endpoint-template.ts](assets/endpoint-template.ts)

---

Hooks

Path:

.claude/settings.json

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "./validate.sh"
          }
        ]
      }
    ],
    "PostToolUse": [...],
    "Notification": [...]
  }
}

Events:

• PreToolUse — before tool execution
• PostToolUse — after execution
• Notification — on notifications

Matcher: tool name or pattern

---

Patterns

Parallel Agents

Run multiple agents simultaneously:

Run in parallel:
1. Task: researcher — study architecture
2. Task: security-scanner — check security
3. Task: performance-analyzer — check performance

Wait for all and combine results.

Progressive Disclosure

Load information as needed:

SKILL.md — brief instructions
references/detailed.md — details when needed
scripts/tool.py — execute without reading into context

Chained Agents

Sequential agent work:

1. researcher → studies the task
2. planner → creates plan based on research
3. implementer → implements the plan
4. reviewer → reviews implementation

---

Validation

Checklist for Commands

• description is filled
• Path: .claude/commands/*.md
• $ARGUMENTS if arguments needed
• Instructions are clear

Checklist for Agents

• name and description are filled
• tools are minimally necessary
• model is chosen consciously
• Path: .claude/agents/*.md

Checklist for Skills

• name is lowercase with hyphens
• description < 1024 characters
• SKILL.md < 500 lines
• Path: .claude/skills/name/SKILL.md

Checklist for Hooks

• JSON is valid
• matcher is correct
• command/script exists
• Path: .claude/settings.json