<p>Complete setup and usage guide for integrating the self-improvement skill with OpenClaw.</p>

<h3>Overview</h3>

<p>OpenClaw uses workspace-based prompt injection combined with event-driven hooks. Context is injected from workspace files at session start, and hooks can trigger on lifecycle events.</p>

<h3>Workspace Structure</h3>

<pre><code>
~/.openclaw/
├── workspace/ # Working directory
│ ├── AGENTS.md # Multi-agent coordination patterns
│ ├── SOUL.md # Behavioral guidelines and personality
│ ├── TOOLS.md # Tool capabilities and gotchas
│ ├── MEMORY.md # Long-term memory (main session only)
│ └── memory/ # Daily memory files
│ └── YYYY-MM-DD.md
├── skills/ # Installed skills
│ └── <skill-name>/
│ └── SKILL.md
└── hooks/ # Custom hooks
└── <hook-name>/
├── HOOK.md
└── handler.ts
</code></pre>

<h3>Quick Setup</h3>

<h4>1. Install the Skill</h4>

<pre><code>
clawdhub install self-improving-agent
</code></pre>

<p>Or copy manually:</p>

<pre><code>
cp -r self-improving-agent ~/.openclaw/skills/
</code></pre>

<h4>2. Install the Hook (Optional)</h4>

<p>Copy the hook to OpenClaw's hooks directory:</p>

<pre><code>
cp -r hooks/openclaw ~/.openclaw/hooks/self-improvement
</code></pre>

<p>Enable the hook:</p>

<pre><code>
openclaw hooks enable self-improvement
</code></pre>

<h4>3. Create Learning Files</h4>

<p>Create the <code>.learnings/</code> directory in your workspace:</p>

<pre><code>
mkdir -p ~/.openclaw/workspace/.learnings
</code></pre>

<p>Or in the skill directory:</p>

<pre><code>
mkdir -p ~/.openclaw/skills/self-improving-agent/.learnings
</code></pre>

<h3>Injected Prompt Files</h3>

<h4>AGENTS.md</h4>

<p>Purpose: Multi-agent workflows and delegation patterns.</p>

<pre><code>

Agent Coordination

Delegation Rules

• Use explore agent for open-ended codebase questions
• Spawn sub-agents for long-running tasks
• Use sessions_send for cross-session communication

Session Handoff

When delegating to another session:

1. Provide full context in the handoff message
2. Include relevant file paths
3. Specify expected output format
   </code></pre>

<h4>SOUL.md</h4>

<p>Purpose: Behavioral guidelines and communication style.</p>

<pre><code>

Behavioral Guidelines

Communication Style

• Be direct and concise
• Avoid unnecessary caveats and disclaimers
• Use technical language appropriate to context

Error Handling

• Admit mistakes promptly
• Provide corrected information immediately
• Log significant errors to learnings
  </code></pre>

<h4>TOOLS.md</h4>

<p>Purpose: Tool capabilities, integration gotchas, local configuration.</p>

<pre><code>

Tool Knowledge

Self-Improvement Skill

Log learnings to .learnings/ for continuous improvement.

Local Tools

• Document tool-specific gotchas here
• Note authentication requirements
• Track integration quirks
  </code></pre>

<h3>Learning Workflow</h3>

<h4>Capturing Learnings</h4>

<ol>
<li><strong>In-session</strong>: Log to <code>.learnings/</code> as usual</li>
<li><strong>Cross-session</strong>: Promote to workspace files</li>
</ol>

<h4>Promotion Decision Tree</h4>

<pre><code>
Is the learning project-specific?
├── Yes → Keep in .learnings/
└── No → Is it behavioral/style-related?
├── Yes → Promote to SOUL.md
└── No → Is it tool-related?
├── Yes → Promote to TOOLS.md
└── No → Promote to AGENTS.md (workflow)
</code></pre>

<h4>Promotion Format Examples</h4>

<p><strong>From learning:</strong></p>
<blockquote>
<p>Git push to GitHub fails without auth configured - triggers desktop prompt</p>
</blockquote>

<p><strong>To TOOLS.md:</strong></p>
<pre><code>

Git

• Don't push without confirming auth is configured
• Use gh auth status to check GitHub CLI a

你在实际生产中遇到过类似问题吗？欢迎留言讨论。
</code></pre>