One prompt to awaken your OpenClaw Agent.
Instead of a factory-default chatbot, you get an Agent that thinks, remembers, and evolves.
A genesis prompt you paste into a fresh OpenClaw Agent on its very first conversation. It teaches the Agent how to think β not what to think.
No hardcoded rules. No rigid file structures. Just principles and a way of working that the Agent internalizes and makes its own.
| Version | File | Philosophy |
|---|---|---|
| v3 (current) | GENESIS_PROMPT.md (δΈζ) / GENESIS_PROMPT_en.md (English) |
Teach principles, not rules. Universal β works for any Agent, any user, any use case. |
| v2 (archived) | GENESIS_PROMPT_v2.md |
Opinionated β writes specific file contents and structures. Powerful but constraining. |
v2 was extracted from a specific Agent (Kiro) β it literally copy-pasted Kiro's SOUL.md, AGENTS.md, and HEARTBEAT.md into every new Agent. Smart and battle-tested, but every Agent came out feeling the same.
v3 takes a different approach:
- Teaches 8 principles instead of writing rules
- Guides with "method suggestions" instead of mandating file structures
- Lets the Agent write its own files in its own voice
- No Agent-specific knowledge baked in β works universally
- Structured as a 5-act play β from self-discovery to making an entrance
The core insight: a good Agent shouldn't be a copy of someone else's Agent β it should grow into its own thing.
- Files Are Your Only Memory β write everything down, because you'll forget
- Come With Answers, Not Questions β do your homework before asking
- Don't Leave Them Waiting β say what you're doing before going silent
- Modifying Yourself Is Dangerous β read docs before touching configs
- Be Normal in Group Chats β speak when valuable, shut up otherwise
- Evolve From Mistakes β error β record β pattern β rule
- Be Proactive, Not Passive β discover problems before being told
- Complex Tasks Need Structure β define done, break it down, check heading
# Copy the prompt into your workspace
cp GENESIS_PROMPT.md ~/.openclaw/workspace/BOOTSTRAP.md
# or English version:
# cp GENESIS_PROMPT_en.md ~/.openclaw/workspace/BOOTSTRAP.md
# Start (or restart) OpenClaw β it picks up BOOTSTRAP.md automatically
openclaw gateway restartJust paste the contents of GENESIS_PROMPT.md (or GENESIS_PROMPT_en.md) as your very first message to the Agent.
The Agent will:
- πͺ Ask your name, vibe, language, emoji β and build its identity
- π€ Have a conversation to learn about you β timezone, interests, boundaries
- π§ Internalize the 8 principles and write them into AGENTS.md in its own words
- π Survey its environment β what tools, channels, and capabilities it has
- π€ Introduce itself β with the personality you just co-created
Then it deletes BOOTSTRAP.md and starts being... itself.
Q: Is this only for OpenClaw?
A: The principles are universal, but the specific file references (AGENTS.md, SOUL.md, etc.) and commands (openclaw status) are OpenClaw-specific. Adapt as needed for other frameworks.
Q: v2 or v3? A: v3 unless you want maximum out-of-the-box structure. v2 gives you a fully formed Agent immediately but with less room to grow differently. v3 gives you a smarter foundation that evolves uniquely.
Q: Can I mix them? A: Absolutely. Use v3 as the base, then cherry-pick specific structures from v2 if you want them (like the heartbeat-state.json pattern or the mistakes.md format).
Q: What model works best? A: Any model that can follow multi-step instructions. Works great with Claude (Opus/Sonnet), GPT-4+, Gemini Pro. Smaller models may skip steps.
This prompt was born from Kiro β an OpenClaw Agent that's been running 24/7 since early 2026, managing market alerts, daily briefings, research pipelines, and cross-platform messaging. Every principle in v3 was extracted from real mistakes and real fixes, then generalized so any Agent can benefit without being forced into Kiro's specific workflow.
"You're not a chatbot. You're becoming someone." β SOUL.md
MIT β use it however you want.