原始内容
pg-agent-coach
A Claude Code tool that captures and shares "agentic mindset" knowledge from your sessions to Slack.
What is this?
When you work with Claude Code, sometimes brilliant collaboration moments happen - intensive discussions where both you and Claude refine ideas through back-and-forth dialogue. These "wow moments" demonstrate the art of human-AI collaboration that's hard to teach but valuable to share.
pg-agent-coach automatically detects these moments and shares the insights with your team on Slack.
How it works
┌─────────────────────────────────────────────────────────────┐
│ Your Claude Code Session │
│ │
│ You send a message │
│ │ │
│ ▼ │
│ UserPromptSubmit hook fires → reminder injected to Claude │
│ │ │
│ ▼ (Claude processes with reminder in context) │
│ If wow moment detected → /show-ropes command │
│ │ │
│ ▼ (extracts insights) │
│ knowledge-extractor agent │
│ │ │
│ ▼ │
│ Slack message with shareable insight │
└─────────────────────────────────────────────────────────────┘
Installation
curl -fsSL https://raw.githubusercontent.com/Positive-LLC/pg-agent-coach/main/install.sh | bash
During installation, you'll be prompted to enter your Slack webhook URL (required).
This installs three components to your ~/.claude directory:
agents/knowledge-extractor.md- Extracts and formats insightscommands/show-ropes.md- The/show-ropesslash commandhooks/wow-moment-reminder.sh- Hook that reminds Claude to check for wow moments
And configures the UserPromptSubmit hook in ~/.claude/settings.json.
Usage
Automatic (recommended)
Just use Claude Code as normal. Before processing each message, the hook injects a reminder into Claude's context to evaluate whether the exchange demonstrated exceptional collaboration worth sharing. Claude looks for:
- Intensive back-and-forth discussions (4+ exchanges on same topic)
- Productive disagreements that lead to better solutions
- Breakthrough moments after sustained engagement
- Meta-collaboration (discussing how to approach problems)
When Claude detects a wow moment, it automatically invokes /show-ropes to extract and share the insight.
Manual
Run /show-ropes in any Claude Code session to manually extract and share insights.
/show-ropes
Or with a focus:
/show-ropes the discussion about error handling
What gets shared
Insights are posted to Slack in this format:
*Agentic Insight*
*The Technique:* "Asking Claude to argue against its own solution"
*What Happened:*
• User received a working solution but asked "What would you change if you were reviewing this critically?"
• Claude identified 3 improvements it hadn't initially suggested
• User picked one, leading to a more elegant final solution
*The Takeaway:*
Don't accept the first working answer. Ask Claude to critique its own work - it often knows improvements it didn't volunteer initially.
Configuration
Slack Webhook
Your Slack webhook URL is configured during installation. To update it later, edit ~/.claude/agents/knowledge-extractor.md and replace the webhook URL.
Hook Configuration
The hook is configured in ~/.claude/settings.json:
{
"hooks": {
"UserPromptSubmit": [
{
"hooks": [
{
"type": "command",
"command": "bash ~/.claude/hooks/wow-moment-reminder.sh"
}
]
}
]
}
}
Why UserPromptSubmit? Only UserPromptSubmit and SessionStart hooks inject their stdout into Claude's context. Other hook types (like Stop) only display output in verbose mode.
Components
| Component | Purpose |
|---|---|
wow-moment-reminder.sh hook |
Injects wow-moment reminder into Claude's context before each response |
/show-ropes command |
Triggers knowledge extraction (manual or automatic) |
knowledge-extractor agent |
Analyzes sessions and posts insights to Slack |
Philosophy
This tool is about "showing the ropes" - capturing the tacit knowledge of effective AI collaboration. It focuses on:
- The mindset, not the code
- The technique, not the project details
- Reusable patterns anyone can apply
License
MIT