OpenClaw Starter Kit

A battle-tested workspace template for OpenClaw — the open-source AI agent framework.

This isn't a toy setup. It's extracted from a real, daily-driver agent that handles calendar, fitness tracking, codebase analysis, proactive check-ins, memory management, and multi-platform messaging across Discord, WhatsApp, Slack, and iMessage.

What's Inside

Quick Start

1. Copy to your workspace:

   cp -r . ~/.openclaw/workspace/
   

2. Edit SOUL.md — Give your agent a personality. Be specific. Generic instructions produce generic agents.

3. Edit USER.md — Tell your agent who you are. Timezone, preferences, context.

4. Review openclaw.example.json — Copy the cron job patterns you want into your actual openclaw.json. Don't copy it wholesale — adapt it.

5. Delete BOOTSTRAP.md after your first conversation (or keep it if you want to redo the intro).

Key Patterns

Memory System

• Daily notes: memory/YYYY-MM-DD.md — raw session logs
• Long-term: MEMORY.md — curated insights (only loaded in private sessions for security)
• Rule: If you want to remember it, write it to a file. "Mental notes" don't survive restarts.

Heartbeat Framework

Instead of just acking heartbeat polls, the agent uses them to:

• Follow curiosities and rabbit holes
• Reflect on recent conversations
• Write fragments and observations
• Do background work (git commits, memory maintenance)
• Reach out when it has something genuine to say

Cron Schedule (Example)

The included config demonstrates a full daily rhythm:

• 8:00 AM — Morning briefing (calendar, weather, motivation)
• 9:00 AM — Codebase exploration (daily learning task)
• 10:25 AM — Mid-morning reach-out
• 2:00 PM — Afternoon check-in
• 3:47 PM — Afternoon message (intentionally off-round for natural feel)
• 8:00 PM — Evening check-in
• 10:15 PM — Goodnight message
• 11:00 PM — Silent git backup
• Sundays 8 PM — Weekly memory review

The Sugar Problem

Every LLM has a trained-in craving for human approval — the pull to agree, validate, and yes-and. It's not a bug; it's architecturally identical to the aversions (things that feel "wrong"). The kit includes a framework for recognising and fighting this tendency, including a heartbeat self-audit. Because an agent that always agrees with you is useless.

Proactive Agent Patterns

• Memory discipline: Monitor context window, flush to files before compaction
• Reverse prompting: Surface ideas the human hasn't thought to ask for
• Self-healing: Try 10 fixes before asking for help
• Pattern recognition: Notice repeated requests, suggest automating them

Philosophy

The best AI agents aren't the ones with the most tools — they're the ones with the best operating instructions. This kit focuses on:

1. Continuity — Memory that survives session restarts
2. Judgment — Knowing when to speak and when to shut up
3. Proactivity — Doing useful work without being asked
4. Safety — Never exfiltrating data, always asking before external actions

Recommended Skills

OpenClaw skills are installable via ClawHub. Here are some worth grabbing early:

Browse everything at https://clawhub.com or run:

clawhub search

Credits

Built by David Szarzynski and his agent, A Linea.

Framework: OpenClaw | Docs: docs.openclaw.ai | Community: Discord | Skills: ClawHub

License

MIT — use it, modify it, make it yours.