Architecture

claude-base is composed of several types of components that work together to help you be more productive.

Overview

File structure

claude-base/
├── .claude/
│   ├── commands/       # <!-- count:commands -->131<!-- /count --> manual commands (/name)
│   │   ├── work/       # Main workflow
│   │   ├── dev/        # Development
│   │   ├── qa/         # Quality
│   │   ├── ops/        # Operations
│   │   ├── doc/        # Documentation
│   │   ├── biz/        # Business
│   │   ├── growth/     # Growth
│   │   ├── data/       # Data
│   │   └── legal/      # Legal
│   ├── agents/         # <!-- count:agents -->63<!-- /count --> autonomous sub-agents
│   ├── skills/         # <!-- count:skills -->54<!-- /count --> auto-triggered skills
│   ├── rules/          # <!-- count:rules -->30<!-- /count --> rules per technology
│   ├── templates/      # Spec/plan templates
│   ├── output-styles/  # Output styles
│   └── settings.json   # Configuration and hooks
├── CLAUDE.md           # Main instructions
└── .mcp.json           # MCP servers configuration

Main components

Commands (131)

Commands are instructions triggered manually with /name.

Characteristics:

• Manual and explicit triggering
• Context shared with the conversation
• All tools available
• .md files in .claude/commands/

Example:

/work:work-explore
/dev:dev-tdd "Implement the user service"
/qa:qa-security

Agents (63)

Agents are autonomous sub-agents with an isolated context.

Characteristics:

• Automatic triggering by delegation
• Isolated context (does not pollute the conversation)
• Restricted tools (read-only for some)
• Specific model (haiku or sonnet)

Delegation example:

"Run a security audit" → Claude delegates to the qa-security agent (sonnet)
"Explore the auth code" → Claude delegates to the work-explore agent (haiku)

Skills (54)

Skills are auto-triggered by keywords in the conversation.

Characteristics:

• Automatic triggering on keywords
• Forked (isolated) or shared context
• Restricted tools via allowed-tools
• SKILL.md files in .claude/skills/

Triggering example:

"I want to do TDD" → test-driven-development skill activated
"Make a commit" → generating-commit-messages skill activated

Rules (30)

Rules are rules applied by file path.

Characteristics:

• Automatic application based on the file
• Specific paths (e.g., **/*.tsx, **/api/**)
• Code conventions per technology

Example:

---
paths:
  - "**/*.ts"
  - "**/*.tsx"
---
# TypeScript rules applied to these files

Comparison

Aspect	Commands	Agents	Skills
Declenchement	Manuel (/nom)	Automatique (delegation)	Automatique (mots-cles)
Contexte	Partage	Isole	Fork ou partage
Modele	Herite du parent	Haiku ou Sonnet	Herite du parent
Outils	Tous disponibles	Restreints	Restreints (allowed-tools)
Cas d'usage	Actions explicites	Taches autonomes	Declenchement contextuel
Nombre	131	63	54

When to use what?

Use a Command when:

• You want an explicit and controlled action
• You need all the tools
• The conversation context is important

Use an Agent when:

• The task can be autonomous
• You want to isolate the context
• The task is standardized (audit, exploration)

Use a Skill when:

• The action is recurring and contextual
• The keywords are specific
• You want automatic triggering

Models used

Model	Usage	Agents
Haiku	Fast, economical tasks	work-explore, doc-onboard, wcag-audit
Sonnet	Complex tasks, analyses	qa-security, qa-audit, dev-debug
Opus	Maximum capabilities	(Not used by default)

Hooks and automations

The .claude/settings.json file configures automatic hooks:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Edit|Write",
        "command": "scripts/validate.sh protect-main"
      }
    ],
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "command": "scripts/validate.sh auto-format $FILE_PATH"
      }
    ]
  }
}

Available hooks:

• Main protection: Blocks modifications on main/master
• Auto-format: Prettier on TS/JS files
• Type-check: TypeScript verification
• Auto-install: npm install after modification of package.json

MCP configuration

The .mcp.json file configures MCP (Model Context Protocol) servers:

{
  "mcpServers": {
    "filesystem": { "enabled": false },
    "memory": { "enabled": false },
    "github": { "enabled": false }
  }
}

Enable the servers as needed to extend Claude's capabilities.

Next steps

• Installation - Complete installation guide
• Workflows - See the workflows in action
• Commands - Explore the 131 commands