Skip to content

agent claude code

Matt Sutton edited this page Mar 14, 2026 · 1 revision

Claude Code

Claude Code serves as the primary depth execution and product management agent. It relies on a blend of project-level files and system directories.

1. Config Surfaces

  • Project Context: CLAUDE.md — The primary source of instructions and project rules.
  • Project Settings: .claude/ — Contains workspace-specific configurations (e.g., settings.local.json).
  • Global Settings: ~/.claude/ — Contains global user preferences.
  • Hooks: Git hooks or tool-specific hooks.
  • MCP Servers: .mcp.json — Defines MCP server commands and environment pass-through.

2. Keeping Configs Lean

Due to the compliance lesson (length kills compliance), CLAUDE.md must only contain critical development rules and the Totem AI Integration block. Avoid turning CLAUDE.md into a massive styleguide. Keep it <32 lines if possible, prioritizing the most important architectural constraints.

3. Totem Integration

Totem injects the search_knowledge instruction (the "Pull Before Coding" reflex) directly into CLAUDE.md. Because Claude uses .mcp.json to discover tools, ensure Totem's MCP server is properly registered there.

4. Common Pitfalls

  • Bloat: Expanding CLAUDE.md beyond 32 lines drastically reduces adherence.
  • Secrets in .mcp.json: Passing hardcoded API keys in .mcp.json. Always use env variables or shell inheritance instead.

Clone this wiki locally