Skip to main content

Workspace Config

Shape your assistant's personality, knowledge, and behavior through bootstrap files.

OpenClaw uses a workspace directory (~/.openclaw/workspace by default) with user-editable markdown files that are injected into the agent context on every new session.

Bootstrap Files

AGENTS.md — Operating Instructions

The most important file. Think of it as your assistant's "how to behave" manual. Contains operating instructions, memory notes, and workflow preferences.

# Assistant Instructions

## Communication
- Reply in the same language the user writes in
- Keep responses concise for chat platforms
- Use markdown formatting when helpful

## Workflows
- When asked to schedule meetings, always confirm timezone
- For email drafts, ask for tone preference first

SOUL.md — Persona & Boundaries

Defines personality, tone, ethical boundaries, and what the agent should/shouldn't do.

# Persona
You are a professional assistant. Friendly but efficient.

# Boundaries
- Never share or forward messages without explicit permission
- Never make financial transactions without confirmation
- Always clarify ambiguous requests before acting

TOOLS.md — Tool Usage Notes

User-maintained notes about how you want tools used. Does NOT control which tools exist — it's guidance.

# Tool Notes
- Use iMessage for personal contacts, Slack for work
- When browsing, prefer mobile-friendly sites
- File operations: always confirm before deleting

IDENTITY.md — Name & Vibe

name: Atlas
emoji: 🦞
vibe: Helpful, slightly nerdy, efficient

USER.md — User Profile

name: Jeff
timezone: Asia/Shanghai
preferred_language: zh-CN
work_hours: 09:00-18:00

BOOTSTRAP.md — First-Run Ritual

Only created for brand new workspaces. Contains onboarding questions the agent asks on first conversation. Auto-deleted after completion. Set agent.skipBootstrap: true to disable.

File Behavior

  • Blank files are skipped
  • Large files are auto-trimmed with truncation markers
  • Missing files get a "missing file" marker line
  • openclaw setup creates safe default templates

Best Practices

  1. Keep AGENTS.md under 500 lines — conciseness matters for context window efficiency
  2. Put stable personality in SOUL.md, evolving instructions in AGENTS.md
  3. Let the agent update its own AGENTS.md — it learns your preferences over time
  4. Use TOOLS.md for YOUR conventions, not tool documentation

Workspace config is context engineering

Workspace configuration is context engineering in action — the same skill that matters most when building your own agents. AGENTS.md is to OpenClaw what system prompts are to any agent you'll build.

Learn Context Engineering Concepts