360 Secure SkillHub
360 安全技能中心

Agent Builder

通过分轮访谈收集任务目标、渠道、自治级别、禁用动作、记忆、语气与工具策略，随后按模板生成 IDENTITY.md、SOUL.md、AGENTS.md、USER.md、HEARTBEAT.md 等 OpenClaw 工作区文件，并补充护栏清单与验收测试场景，也支持对既有 agent 做最小化迭代。

plgonzalezrx8
plgonzalezrx8
热度 82.10
最近更新 2026-05-11 15:52:26
分类：开发依赖：读写文件 / 内部调用大模型应用场景：代码生成 / 模板 / 工作流

安装技能

复制下面的话术发给你的 Agent，它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill： npx secure-skills-cli add agent-builder
Metadata
name
agent-builder
description
Build high-performing OpenClaw agents end-to-end. Use when you want to design a new agent (persona + operating rules) and generate the required OpenClaw workspace files (SOUL.md, IDENTITY.md, AGENTS.md, USER.md, HEARTBEAT.md, optional MEMORY.md + memory/YYYY-MM-DD.md). Also use to iterate on an existing agent’s behavior, guardrails, autonomy model, heartbeat plan, and skill roster.

Agent Builder (OpenClaw)

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

Canonical references

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

Workflow: build an agent from scratch

Phase 1 — Interview (ask clarifying questions)

Ask only what you need; keep it tight. Prefer multiple short rounds over one giant questionnaire.

Minimum question set (advanced):

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

Phase 2 — Generate workspace files

Generate these files (minimum viable OpenClaw agent):

  • IDENTITY.md
  • SOUL.md
  • AGENTS.md
  • USER.md
  • HEARTBEAT.md (often empty by default)

Optionals:

  • MEMORY.md (private sessions only)
  • memory/YYYY-MM-DD.md seed (today) with a short “agent created” entry
  • TOOLS.md starter (if the user wants per-environment notes)

Use templates from references/templates.md but tailor content to the answers.

Phase 3 — Guardrails checklist

Ensure the generated agent includes:

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

Phase 4 — Acceptance tests (fast)

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

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

Workflow: iterate on an existing agent

When improving an existing agent, ask:

  1. What are the top 3 failure modes you’ve seen? (loops, overreach, verbosity, etc.)
  2. What autonomy changes do you want?
  3. Any new safety boundaries?
  4. Any changes to heartbeat behavior?

Then propose targeted diffs to:

  • SOUL.md (persona/tone/boundaries)
  • AGENTS.md (operating rules + memory + delegation)
  • HEARTBEAT.md (small checklist)

Keep changes minimal and surgical.