21 OpenClaw Power User Tricks That Will Change How You Work

#You Installed OpenClaw. Now What?

Most people install OpenClaw, connect it to Telegram or Discord, ask it a few questions, and think they are done. They are using maybe 5% of what the platform can do.

The gap between "I have OpenClaw running" and "OpenClaw runs my life" is not about spending more on API tokens. It is about knowing the tricks that make an agent genuinely autonomous, reliable, and useful.

These are the tricks we use daily. Not theory, real configurations and workflows that work in production.

#Memory & Identity

#1. Write a Real SOUL.md (Not a Job Description)

Most people either skip SOUL.md entirely or write something generic like "You are a helpful assistant." This is a waste.

Your SOUL.md shapes every response your agent gives. Tell it what tone to use. Tell it what you hate. Give it opinions. The more specific you are, the less you have to correct it later.

Bad:

You are a helpful AI assistant.

Good:

Be concise. I hate filler words. Never say "Great question!"
Default to action over discussion. If you can fix it, fix it.
When I ask for options, give me your recommendation first.

The difference in response quality is night and day.

#2. Use AGENTS.md to Set Behavioral Rules

SOUL.md is personality. AGENTS.md is operating procedure. Use it to define things like:

• What your agent should do at the start of every session
• Which files to check automatically
• What actions require permission vs. what it can do freely
• How to handle group chats vs. direct messages

Think of AGENTS.md as the employee handbook and SOUL.md as the personality test.

#3. Build a Memory System That Actually Works

OpenClaw agents wake up fresh every session. If you do not build a memory system, your agent has amnesia.

The pattern that works:

• Daily files (memory/YYYY-MM-DD.md), raw notes about what happened each day
• Long-term memory (MEMORY.md), curated insights, preferences, key decisions
• Session startup: aGENTS.md tells the agent to read today's and yesterday's daily files plus MEMORY.md before doing anything

Your agent should update daily files during conversations and periodically review them to update MEMORY.md. This gives you an agent that actually learns over time.

#4. Use IDENTITY.md for Multi-Channel Consistency

If your agent operates across Telegram, Discord, and email, give it a name and consistent identity in IDENTITY.md. Include a preferred emoji and avatar. This makes it feel like the same entity everywhere instead of a generic bot.

#Automation & Scheduling

#5. Heartbeats Are Not Just Health Checks

Most people think heartbeats are just "is the agent alive?" pings. Wrong. Heartbeat is "a system-owned automation that runs periodic agent turns in the main session so the model can surface anything that needs attention without spamming you" — your agent's standing chance to be proactive, every 30 minutes by default.

The checklist does not go in a HEARTBEAT.md file any more. That file is retired: OpenClaw "no longer creates HEARTBEAT.md in new workspaces or reads it at runtime." Instructions live in the heartbeat job's monitor scratch instead. Find the job, which lists as Heartbeat (agent-id):

openclaw automations list --all

Then load your checklist into its scratch:

openclaw automations scratch <jobId> --set "- Check email for urgent messages
- Check calendar for upcoming events (next 4h)
- Review memory files for anything to update
- If something needs attention, message me"

Your agent evaluates that checklist on every heartbeat turn. Rotate the checks through the day so you are not burning tokens on the same thing every cycle — and note that if scratch is effectively empty (blank lines, comments or headings only) OpenClaw skips the run entirely and logs reason=empty-heartbeat-file, which is a cheap way to park heartbeat without disabling it.

Still have a HEARTBEAT.md lying around? openclaw doctor --fix imports its instructions into monitor scratch, converts valid legacy task entries into cron jobs, archives the file and removes it from the workspace.

#6. Use Cron for Precision, Heartbeats for Batching

Both are automations now — heartbeat is just the system-owned one — but they earn their keep differently. The docs split it as: automations for "Explicit reports, reminders, recurring work" on a "one-shot, interval, or cron expression" schedule, heartbeat for "Ambient monitoring and event follow-up" at scheduler-controlled intervals. Put another way: "Create an automation for work with its own instructions or schedule; use heartbeat as the system-owned ambient monitor when periodic main-session awareness is useful."

Cron jobs run at their own scheduled times. Heartbeat runs periodic turns in your main session and can batch multiple checks together.

Rule of thumb:

• Morning briefing at 8 AM sharp? Cron.
• Check email and calendar together a few times a day? Heartbeat.
• Weekly report every Monday? Cron.
• Quick check if anything changed in the last hour? Heartbeat.

Mixing both gives you precise scheduling AND efficient batch processing without burning unnecessary API calls.

#7. Build a Cron Safety Net

Cron jobs can fail silently. Add a check to your heartbeat scratch that verifies your critical jobs actually ran. Note the run-history syntax: the job id is positional, not a --id flag.

## Cron Safety Net (check between 8-10 AM)
1. Run `openclaw automations runs <job-id> --limit 1`
2. If today's run shows error status, execute the task manually
3. If it hasn't run yet, do the task yourself

The Gateway applies run-history filters before paging, so you can narrow this further: --status takes all, ok, error or skipped, and --delivery-status takes delivered, not-delivered, unknown or not-requested. A tighter version of step 1 is openclaw automations runs <job-id> --status error --limit 1.

This way a cron failure never means a missed briefing.

#8. Let the Scratch Track Your Heartbeat State

You do not need a side-car JSON file to remember what your agent last checked. The monitor scratch is writable from inside the heartbeat turn: the agent can call heartbeat_respond with a scratch value to "fully replace the monitor scratch for future heartbeats". So tell it to keep its own ledger in the checklist it reads:

## Periodic Checks
- Email: last checked 2026-09-19 08:40
- Calendar: last checked 2026-09-19 08:40
- Weather: never
After each pass, rewrite this section with the current timestamps via
heartbeat_respond so the next heartbeat knows what it can skip.

Scratch "is stored in the shared state database, capped at 256 KiB, and never included in automations list/automations get/automations runs output", so the ledger stays private to the job. Use --expected-revision <n> for compare-and-swap protection if you also edit it from the CLI. This prevents duplicate checks and lets your agent space out its proactive work intelligently, without a file that nothing else in OpenClaw knows about.

#Model & Cost Optimization

#9. Use Model Aliases Instead of Full Names

Instead of typing anthropic/claude-opus-4-6 every time, set up aliases in your config:

opus → anthropic/claude-opus-4-6
kimi → kimi-coding/k2p5
codex53 → openai-codex/gpt-5.3-codex

Faster to type, easier to remember, and you can reroute an alias to a different model without changing anything else.

#10. Set Up Model Failover Chains

API rate limits and outages happen. Configure a failover chain so your agent automatically tries the next model:

Primary: opus
Fallback: opus45 → kimi → codex53 → gpt52

Your agent stays responsive even when your primary provider is having a bad day.

#11. Route Expensive Tasks to Cheaper Models

Not every task needs your most expensive model. Use sub-agents with specific model overrides for different work:

• Research and summarization → Gemini Flash or GPT-4o Mini (cheap, fast, good enough)
• Coding → Codex or Claude Code (specialized, worth the cost)
• Complex reasoning → Opus (expensive but necessary)
• Routine automation → Whatever is cheapest

One Opus call costs more than 50 Flash calls. Route accordingly.

#12. Monitor Your Spending with /status

Get in the habit of checking /status or running session_status to see your token usage and costs. Most people have no idea how much they are spending until the bill arrives.

Set a mental budget. If you are blowing through it, look at which tasks are eating tokens and consider routing them to cheaper models.

#Skills & Integrations

#13. Check ClawHub Before Building From Scratch

Before you spend an hour writing a custom automation, check ClawHub. There are thousands of community skills covering everything from email management to smart home control. Installing a skill takes seconds:

openclaw skills install @<publisher>/<slug>

Even if a skill does not do exactly what you want, it is usually faster to modify an existing one than to start from zero.

14. Keep Tool-Specific Notes in the ## Tools Section of AGENTS.md

TOOLS.md is retired. "Local tool and environment notes now live in the ## Tools section of AGENTS.md", and openclaw doctor --fix will "archive an existing workspace TOOLS.md, merge customized content into AGENTS.md, and remove the retired file."

The cheat sheet itself is unchanged, just relocated:

• Camera names and locations
• SSH hosts and connection details
• Preferred TTS voices
• Account IDs and which calendar is which
• Device nicknames

Skills define how tools work. The ## Tools section tells your agent your specific setup. Keep them separate so you can update skills without losing your configuration.

Two caveats. First, this section is guidance, not wiring: "The ## Tools section holds local environment notes and conventions. It does not control tool availability; it is only guidance." Second, keep credentials out of it — AGENTS.md is a plaintext file the agent reads into context every session. API keys belong in ~/.openclaw/secrets.json, in the global .env at ~/.openclaw/.env, or in the config's secret references, not in a workspace markdown file.

#15. Use Sub-Agents for Heavy Lifting

Do not make your main agent do everything. Spawn sub-agents for tasks that:

• Take a long time (coding, research, document generation)
• Need a different model (use a coding-specialized model for code)
• Should run in isolation (so failures do not crash your main session)
• Can run in parallel (research three topics simultaneously)

Sub-agents report back when done. Your main agent stays responsive for conversation while work happens in the background.

#Communication & Channels

#16. Set Up Platform-Specific Formatting Rules

Different platforms render content differently. Put this in your AGENTS.md:

## Platform Formatting
- Discord/WhatsApp: No markdown tables — use bullet lists
- Discord: Wrap links in <> to suppress embeds
- WhatsApp: No headers — use **bold** for emphasis
- Telegram: Full markdown supported

This prevents your agent from sending garbled messages across channels.

#17. Master the Silent Reply

When your agent has nothing useful to add, it should say nothing. Configure the NO_REPLY behavior so your agent does not spam group chats with "Thanks!" and "Got it!" messages.

In group settings, your agent should only speak when:

• Directly mentioned or asked a question
• It can add genuine value
• Something witty fits naturally
• Correcting important misinformation

Silence is a feature, not a bug.

#18. Use Reactions Instead of Replies

On platforms that support emoji reactions (Discord, Telegram, Slack), configure your agent to react instead of replying when a full message is not needed. A 👍 says "I acknowledge this" without cluttering the conversation.

One reaction is always better than a "Sounds good!" reply that nobody needed.

#Security & Maintenance

#19. Never Expose Your Gateway to the Public Internet

This should be obvious but apparently it is not, over 135,000 OpenClaw instances are currently exposed to the public internet. 93% have authentication bypass vulnerabilities.

Your OpenClaw gateway should only be accessible from localhost or your local network. If you need remote access, use a VPN or SSH tunnel. Never, ever open port 3000 to the world.

#20. Audit Your Installed Skills Regularly

Over 824 confirmed malicious skills have been found in OpenClaw skill registries. Before installing any skill:

• Check the source repository
• Read the SKILL.md and any scripts it runs
• Look at what permissions it requires
• Prefer skills from known/verified publishers

Run openclaw skills list periodically and remove anything you are not actively using. Every installed skill is attack surface.

#21. Use trash Instead of rm

Configure your agent to use trash instead of rm for file deletion. Recoverable beats gone forever. One wrong rm -rf and you are having a very bad day. Put this in AGENTS.md as a hard rule.

#The Meta-Trick: Document Everything

The single highest-leverage thing you can do with OpenClaw is document your workflows. Every time you figure out a trick, solve a problem, or establish a preference, write it down in your workspace files.

Your agent reads these files at the start of every session. Every documented decision is a decision you never have to make again. Every noted preference is a correction you never have to give again.

The people who get the most out of OpenClaw are not the ones with the fanciest setups. They are the ones who methodically teach their agent how they work.

Start with SOUL.md and AGENTS.md, including its ## Tools section. Add MEMORY.md. Build daily memory files. Before you know it, you have an agent that actually knows you.

#Want to Go Deeper?

These tricks scratch the surface. For complete SOUL.md templates, pre-built automation skills, and step-by-step guides for every workflow mentioned here, check out the ClawDocx library. We build and test every resource with real OpenClaw setups, not theory, practice.