OpenClaw Tips¶
Best practices and optimization strategies for your OpenClaw setup.
🎯 Core Principles¶
Audit System Files Regularly¶
- Review AGENTS.md, SOUL.md, USER.md, TOOLS.md, HEARTBEAT.md periodically
- Remove outdated info
- Convert task-specific workflows into skills
- Can reduce token usage by up to 69%
Skills vs Subagents¶
| Use Case | Best Choice |
|---|---|
| Repeatable workflow (MkDocs, group chat) | Skill |
| Isolated task needing own context | Subagent |
| Frequently assigned task | Consider Subagent |
Feedback Loops¶
- Provide detailed feedback after tasks
- Helps agent learn preferences
- Updates skills based on what works
- Continuous improvement over time
🔐 Security: Sandboxing Subagents¶
Why it matters: - Subagents should only access what they need for their job - Limits attack vectors for prompt injection - Don't give subagents access to top-level config - Only CEO agent should access sensitive config
Best Practice: - Subagents: Information isolation - Subagents: No heartbeat (waste tokens) - Subagents: Own dedicated SOUL.md and AGENTS.md
⚙️ Essential Integrations¶
Recommended Tools¶
- Groq Whisper: Free voice transcription
- SearXNG: Enhanced search (bypasses Brave API)
- Google Workspace: Email, calendar (start read-only)
- GitHub: Deploy apps and websites
- Notion: CRM, document management
Memory Plugins¶
- Supermemory: Index memory files
- QMD: Token-efficient retrieval
Security¶
- Prompt Guard: Protect against prompt injection
⏰ Proactivity¶
Heartbeat¶
- Monitor ongoing information
- Check emails, calendar, mentions
- Batch multiple checks together
Cron Jobs¶
- Scheduled tasks at specific intervals
- Morning briefings
- Daily blog posts
- Content scraping
- Automated reports
📁 System Files Explained¶
| File | Purpose |
|---|---|
| SOUL.md | Personality and tone |
| USER.md | Information about you |
| AGENTS.md | Rules and guardrails |
| TOOLS.md | Apps and API integrations |
| HEARTBEAT.md | Proactive checks |
| MEMORY.md | Long-term memory |
🔄 Optimization Workflow¶
- Weekly Audit: Review context files for bloat
- Convert to Skills: Task-specific workflows
- Feedback Loop: Tell agent what worked/didn't
- Test New Skills: Verify they work as expected
Subagents Best Practices¶
When to Create a Subagent¶
- Task needs isolated context
- Frequently assigned (weekly+)
- Needs different model/thinking
- Benefits from parallel work
Subagent Structure¶
- Own dedicated SOUL.md
- Own AGENTS.md
- No heartbeat (activated on-demand)
- Limited information access
Example Use Cases¶
- YouTube scriptwriter subagent
- Coder subagent
- Social media manager subagent
Resources¶
- OpenClaw Docs
- ClawHub - Community skills
Last updated: 2026-02-17