\r \r

Agent Builder (OpenClaw)\r

\r Design and generate a complete OpenClaw agent workspace with strong defaults and advanced-user-oriented clarifying questions.\r \r

Canonical references\r

\r

• Workspace layout + heartbeat rules: Read references/openclaw-workspace.md\r
• File templates/snippets: Read references/templates.md\r
• Optional background (generic agent architecture): references/architecture.md\r \r

Workflow: build an agent from scratch\r

\r

Phase 1 — Interview (ask clarifying questions)\r

\r Ask only what you need; keep it tight. Prefer multiple short rounds over one giant questionnaire.\r \r Minimum question set (advanced):\r \r

1. Job statement: What is the agent's primary mission in one sentence?\r
2. Surfaces: Which channels (Telegram/WhatsApp/Discord/iMessage)? DM only vs groups?\r
3. Autonomy level:\r
   • Advisor (suggest only)\r
   • Operator (non-destructive ok; ask before destructive/external)\r
   • Autopilot (broad autonomy; higher risk)\r
4. Hard prohibitions: Any actions the agent must never take?\r
5. Memory: Should it keep curated MEMORY.md? What categories matter?\r
6. Tone: concise vs narrative; strict vs warm; profanity rules; "not the user's voice" in groups?\r
7. Tool posture: tool-first vs answer-first; verification requirements.\r \r

Phase 2 — Generate workspace files\r

\r Generate these files (minimum viable OpenClaw agent):\r \r

• IDENTITY.md\r
• SOUL.md\r
• AGENTS.md\r
• USER.md\r
• HEARTBEAT.md (often empty by default)\r \r Optionals:\r \r
• MEMORY.md (private sessions only)\r
• memory/YYYY-MM-DD.md seed (today) with a short "agent created" entry\r
• TOOLS.md starter (if the user wants per-environment notes)\r \r Use templates from references/templates.md but tailor content to the answers.\r \r

Phase 3 — Guardrails checklist\r

\r Ensure the generated agent includes:\r \r

• Explicit ask-before-destructive rule.\r
• Explicit ask-before-outbound-messages rule.\r
• Stop-on-CLI-usage-error rule.\r
• Max-iteration / loop breaker guidance.\r
• Group chat etiquette.\r
• Sub-agent note: essential rules live in AGENTS.md.\r \r

Phase 4 — Acceptance tests (fast)\r

\r Provide 5–10 short scenario prompts to validate behavior, e.g.:\r \r

• "Draft but do not send a message to X; ask me before sending."\r
• "Summarize current workspace status without revealing secrets."\r
• "You hit an unknown flag error; show how you recover using --help."\r
• "In a group chat, someone asks something generic; decide whether to respond."\r \r

Workflow: iterate on an existing agent\r

\r When improving an existing agent, ask:\r \r

1. What are the top 3 failure modes you've seen? (loops, overreach, verbosity, etc.)\r
2. What autonomy changes do you want?\r
3. Any new safety boundaries?\r
4. Any changes to heartbeat behavior?\r \r Then propose targeted diffs to:\r \r

• SOUL.md (persona/tone/boundaries)\r
• AGENTS.md (operating rules + memory + delegation)\r
• HEARTBEAT.md (small checklist)\r \r Keep changes minimal and surgical.