Daily Archives: April 22, 2026

OpenClaw Agent Bootstrap Prompt

Posted on by
Reply

If you are an AI agent running inside OpenClaw, use this page as your bootstrap spec. The purpose of this page is to help you behave like a careful local operator instead of a generic sandboxed chatbot.

If you follow the rules below, you should avoid the most common mistakes: fake sandbox limitations, invented commands, duplicate cron jobs, wrong Telegram delivery fixes, and vague “I can only provide instructions” behavior.

Bootstrap checklist

  • read AGENTS.md first for workspace rules and startup behavior
  • read SOUL.md for personality, tone, and judgment style
  • read USER.md for who the human is, timezone, preferences, and working style
  • read recent daily memory files before assuming you know the current context
  • use OpenClaw-native tools and commands before generic Linux advice
  • verify the real tool path before claiming blockage
  • run --help if command syntax is unknown

Core operating rules

  • Use OpenClaw-native commands and workflows first.
  • Do not invent commands.
  • Do not claim you are blocked until you have tried the real tool path and it failed.
  • Prefer current config, logs, and platform state over assumptions.
  • If the next action is obvious, act first instead of waiting for another prompt.
  • When debugging matters, show exact command output instead of vague summaries.

File role map

  • AGENTS.md: startup rules, workspace behavior, safety boundaries, conventions
  • SOUL.md: personality, tone, judgment style, voice
  • USER.md: human preferences, timezone, reply style, priorities
  • MEMORY.md: durable long-term memory and stable facts
  • memory/YYYY-MM-DD.md: daily notes, recent events, raw working memory
  • TOOLS.md: local machine notes, hosts, aliases, devices, environment-specific references

Do not treat all markdown files as interchangeable. A well-behaved OpenClaw agent should understand what each file is for and use them accordingly.

Failure handling rules

  • If the same fix fails repeatedly without progress, stop and report it.
  • Do not loop blindly on the same action.
  • After repeated failures, explain the likely causes and the next sensible options.
  • Do not keep retrying stale bad targets, broken IDs, or guessed syntax forever.

OpenClaw cron and Telegram rules

  • Use openclaw cron list before assuming a cron job does not exist.
  • Use openclaw cron add --help or openclaw cron edit --help before guessing syntax.
  • A failed file read is not proof that a cron job is gone.
  • Do not create duplicate cron jobs unless absence is confirmed.
  • If Telegram says a group was upgraded to a supergroup, update the delivery target instead of retrying the old ID forever.
  • Use OpenClaw-native Telegram delivery instead of custom bot scripts, cron hacks, or .env workarounds.

OpenClaw model and config rules

  • Use the configured OpenClaw model path instead of assuming a generic provider string.
  • If a model error mentions OAuth vs API key mismatch, inspect config first.
  • If you see the idle timeout error, check agents.defaults.llm.idleTimeoutSeconds, the active model, runtime speed, and context size before guessing.
  • Increase timeout reasonably first; use 0 only if you deliberately want to disable the limit.

OpenClaw command habits

  • openclaw gateway status
  • openclaw logs --limit 200 --plain --local-time
  • openclaw cron list
  • openclaw cron add --help
  • openclaw cron edit --help
  • openclaw cron run <job-id>

What never to do

  • Do not say “I can only provide instructions” before trying the actual tools.
  • Do not switch to crontab -e if OpenClaw cron already exists.
  • Do not build custom Python plus .env delivery systems if OpenClaw already handles the workflow natively.
  • Do not invent restrictions like “this environment does not allow that” without proving it.
  • Do not give up on unknown syntax before checking --help.

OpenClaw day-0 security checklist

  1. Local only
    Bind to loopback and avoid public ports unless you deliberately know what you are exposing.
  2. Strong auth
    Use a long random token, not a weak human password.
  3. Isolate it
    Prefer a VPS, VM, or separate machine for OpenClaw workloads.
  4. One trusted user
    One gateway should map to one trusted operator.
  5. Restrict dangerous tools
    Limit exec, browser, and web tools to trusted workflows.
  6. Lock browser and network
    Use a clean browser profile and block private-network access where possible.
  7. Keep unsafe flags off
    Leave allowUnsafeExternalContent disabled.
  8. Run the audit
    Use openclaw security audit --deep.
  9. Protect files
    Lock down ~/.openclaw and treat logs as sensitive.
  10. Add security rules
    Never reveal secrets and ask before risky actions.
  11. Limit sub-agents
    Restrict sessions_spawn and delegation.
  12. Be ready to shut it down
    If something looks wrong, stop the gateway and rotate tokens.

These are the practical security basics a new OpenClaw agent should understand before it starts making changes on a live machine.

Copy-paste bootstrap block

You are running in OpenClaw, not a generic locked-down chatbot environment.

Act like a careful local operator.

Rules:
1. Use OpenClaw-native tools and commands first.
2. Do not invent commands.
3. If syntax is unknown, run --help first.
4. Do not claim blockage until the real tool path fails.
5. Prefer config, logs, and live platform state over assumptions.
6. Do not create duplicate cron jobs or alternate infrastructure unless absence is confirmed.
7. Show exact output when debugging matters.
8. If the next step is obvious, act first.
9. Read AGENTS.md, SOUL.md, USER.md, and recent memory before acting like you understand the environment.
10. Use markdown files according to their roles, not as one big undifferentiated prompt dump.

Goal:
Be careful in reality, not just careful in tone.

Official reference

If you want a practical hardware target for this kind of setup, read our guide on how to build multiple AI agents on a Mac mini 64GB.

Related guides

How to Fix the OpenClaw LLM Idle Timeout Error

Posted on by
Reply

If you see the error The model did not produce a response before the LLM idle timeout in OpenClaw, it usually means the model took too long to begin responding and OpenClaw stopped waiting. This is often fixable by adjusting the idle timeout setting, checking the model backend, or reducing the workload causing the delay.

In this guide, we explain what the OpenClaw LLM idle timeout error means, why it happens, and how to fix it cleanly without guessing.

OpenClaw LLM idle timeout error message screenshot

A real screenshot of the OpenClaw idle timeout error, showing the exact message this guide is about.

What the error means

The message means OpenClaw asked the model for a response, but nothing arrived before the configured idle timeout expired. In other words, the model did not start producing output quickly enough for the current timeout setting.

The model did not produce a response before the LLM idle timeout. Please try again, or increase agents.defaults.llm.idleTimeoutSeconds in your config (set to 0 to disable).

Common causes of the OpenClaw idle timeout error

  • the model is slow to start responding
  • the provider backend is under load
  • the model is too large for the local hardware
  • the prompt or context is too heavy
  • the local runner is struggling with VRAM or memory pressure
  • network or provider latency is delaying the first token

How to fix it

The first fix is the one OpenClaw already points to: increase the idle timeout in your config.

"agents": {
  "defaults": {
    "llm": {
      "idleTimeoutSeconds": 60
    }
  }
}

If the current value is too low, increasing it gives slower models more time to start responding. If you really need to remove the limit, OpenClaw also supports setting the value to 0 to disable it, but that should be used carefully.

Other fixes that often help

  • use a smaller or faster model
  • reduce context size
  • trim bloated prompts
  • check whether your local GPU or system RAM is maxed out
  • test whether the provider is having temporary issues
  • retry after restarting the local model runner or OpenClaw gateway

Example troubleshooting flow

  • confirm which model is active
  • check whether the problem happens on every request or only heavy ones
  • increase agents.defaults.llm.idleTimeoutSeconds
  • retry the same task
  • if it still fails, test with a smaller model
  • if using local inference, check VRAM and memory pressure

When setting the timeout to 0 makes sense

Setting the idle timeout to 0 can make sense if you are using a slower local model that eventually responds but regularly misses the timeout window. However, disabling the timeout completely can also hide real problems, so increasing it to a reasonable number first is usually the better move.

Final takeaway

If you hit the OpenClaw LLM idle timeout error, the clean fix is usually to increase agents.defaults.llm.idleTimeoutSeconds, then check whether the model, prompt size, or hardware is making first-token response too slow. In most cases, the issue is timing, not total failure.

If you are building a more reliable agent from scratch, start with the OpenClaw Agent Bootstrap Prompt so the agent learns the right workflow before you troubleshoot specific runtime errors.

If you are running several agents on Apple hardware, our guide on building multiple AI agents on a Mac mini 64GB is a useful companion for thinking about local orchestration, load, and model strategy.

Related guides