A Claude Code skill for configuring, diagnosing, and training Openclaw agents.
When you have questions about Openclaw — why an agent isn't responding, how to write a good AGENTS.md, how to build a Hook or Skill — this skill reads the relevant knowledge docs before answering, instead of guessing from memory.
git clone https://github.com/weilun88313/openclaw-coach-skill ~/.claude/skills/openclaw-coachThat's it. The skill auto-loads in every Claude Code session.
Covers four scenarios:
| Scenario | Examples |
|---|---|
| Diagnose | Agent not responding, agent forgets conversations, three agents behave identically |
| Configure | Write or improve AGENTS.md, SOUL.md, USER.md, MEMORY.md, HEARTBEAT.md |
| Develop | Build a Skill, Hook, or MCP integration |
| Advanced | Set up autonomous task mode, Heartbeat/Cron, sessions_spawn sub-agents |
Core principle: the skill always reads the relevant knowledge doc before giving advice. It won't guess.
Docs live in knowledge/:
| File | Covers |
|---|---|
Openclaw-入门.md |
Quickstart, channel setup (WhatsApp/Telegram/Discord), multi-agent routing, common errors |
Openclaw-核心-文件与记忆.md |
Workspace file roles, size limits, three-layer memory model, write rules, file templates |
Openclaw-核心-搜索与多Agent.md |
Agent Reach install, search decision tree, plan-file mode, SHARED.md shared memory |
Openclaw-三件套-Skill.md |
Skill development: trigger design, SKILL.md format, soft-trigger patterns |
Openclaw-三件套-Hook.md |
Hook development: event list, handler.ts template, compact-guardian pattern |
Openclaw-三件套-MCP.md |
MCP integration: mcporter config, server ecosystem, when to use MCP vs Hook |
Openclaw-进阶-自主任务.md |
Autonomous mode: AUTONOMOUS.md, Heartbeat vs Cron, sub-agent tool limits |
To add a new doc: drop it in knowledge/, add a row to the document map in SKILL.md.
Three test cases in evals/evals.json:
| # | Scenario | Flow tested |
|---|---|---|
| 1 | Agent forgets previous conversations | Diagnose → memory model |
| 2 | Write an AGENTS.md for a work assistant agent |
Configure → workspace files |
| 3 | Auto-restore task context after compact |
Develop → Hook (compact-guardian) |
Run evals with the skill-creator skill.
The skill activates on mentions of:
openclaw · agent 配置 · 训练 agent · AGENTS.md · SOUL.md · HEARTBEAT.md · MEMORY.md · agent 不回复 · agent 笨 · agent 失忆 · Hook 开发 · Skill 开发 · MCP 集成 · 三件套 · compact · Heartbeat · 自主任务 · sessions_spawn
Knowledge gaps are tracked in the "知识库缺口处理" section of SKILL.md. If you find a scenario the skill handles poorly, open an issue describing the prompt and what the skill got wrong.
PRs that add or improve knowledge docs are welcome. Keep each doc under 300 lines; add a table of contents if longer.
MIT