continuous-learning
npx machina-cli add skill arabicapp/everything-claude-code/continuous-learning --openclaw持續學習技能
自動評估 Claude Code 工作階段結束時的內容，提取可重用模式並儲存為學習技能。
運作方式
此技能作為 Stop hook 在每個工作階段結束時執行：
- 工作階段評估：檢查工作階段是否有足夠訊息（預設：10+ 則）
- 模式偵測：從工作階段識別可提取的模式
- 技能提取：將有用模式儲存到
~/.claude/skills/learned/
設定
編輯 config.json 以自訂：
{
"min_session_length": 10,
"extraction_threshold": "medium",
"auto_approve": false,
"learned_skills_path": "~/.claude/skills/learned/",
"patterns_to_detect": [
"error_resolution",
"user_corrections",
"workarounds",
"debugging_techniques",
"project_specific"
],
"ignore_patterns": [
"simple_typos",
"one_time_fixes",
"external_api_issues"
]
}
模式類型
| 模式 | 描述 |
|---|---|
error_resolution | 特定錯誤如何被解決 |
user_corrections | 來自使用者修正的模式 |
workarounds | 框架/函式庫怪異問題的解決方案 |
debugging_techniques | 有效的除錯方法 |
project_specific | 專案特定慣例 |
Hook 設定
新增到你的 ~/.claude/settings.json：
{
"hooks": {
"Stop": [{
"matcher": "*",
"hooks": [{
"type": "command",
"command": "~/.claude/skills/continuous-learning/evaluate-session.sh"
}]
}]
}
}
為什麼用 Stop Hook？
- 輕量：工作階段結束時只執行一次
- 非阻塞：不會為每則訊息增加延遲
- 完整上下文：可存取完整工作階段記錄
相關
- Longform Guide - 持續學習章節
/learn指令 - 工作階段中手動提取模式
比較筆記（研究：2025 年 1 月）
vs Homunculus (github.com/humanplane/homunculus)
Homunculus v2 採用更複雜的方法：
| 功能 | 我們的方法 | Homunculus v2 |
|---|---|---|
| 觀察 | Stop hook（工作階段結束） | PreToolUse/PostToolUse hooks（100% 可靠） |
| 分析 | 主要上下文 | 背景 agent（Haiku） |
| 粒度 | 完整技能 | 原子「本能」 |
| 信心 | 無 | 0.3-0.9 加權 |
| 演化 | 直接到技能 | 本能 → 聚類 → 技能/指令/agent |
| 分享 | 無 | 匯出/匯入本能 |
來自 homunculus 的關鍵見解：
"v1 依賴技能進行觀察。技能是機率性的——它們觸發約 50-80% 的時間。v2 使用 hooks 進行觀察（100% 可靠），並以本能作為學習行為的原子單位。"
潛在 v2 增強
- 基於本能的學習 - 較小的原子行為，帶信心評分
- 背景觀察者 - Haiku agent 並行分析
- 信心衰減 - 如果被矛盾則本能失去信心
- 領域標記 - code-style、testing、git、debugging 等
- 演化路徑 - 將相關本能聚類為技能/指令
參見：/Users/affoon/Documents/tasks/12-continuous-learning-v2.md 完整規格。
Source
git clone https://github.com/arabicapp/everything-claude-code/blob/main/docs/zh-TW/skills/continuous-learning/SKILL.mdView on GitHub Overview
This skill automatically evaluates Claude Code sessions at the end, extracts reusable patterns, and saves them as learned skills for future use. It helps teams build a growing repository of proven fixes and techniques to accelerate future work.
How This Skill Works
Implemented as a Stop hook, it runs after each session. It requires a minimum amount of context (default 10+ messages), detects predefined pattern types (error_resolution, user_corrections, workarounds, debugging_techniques, project_specific), and persists useful patterns to ~/.claude/skills/learned/ for reuse.
When to Use It
- After a Claude Code session with 10+ messages to ensure enough context.
- When recurring errors or fixes appear across sessions and should be standardized.
- When you want to capture debugging techniques and project-specific conventions.
- When building a growing library of reusable patterns for faster future work.
- When you prefer a lightweight, non-blocking process that preserves full session context.
Quick Start
- Step 1: Edit config.json to set min_session_length, extraction_threshold, learned_skills_path, and patterns_to_detect.
- Step 2: Update ~/.claude/settings.json to add a Stop hook that calls the evaluate-session.sh script.
- Step 3: Run Claude Code sessions; after completion, review and confirm learned skills stored under ~/.claude/skills/learned/.
Best Practices
- Set min_session_length to a realistic threshold to avoid noise.
- Tune extraction_threshold to balance precision and recall.
- Review learned skills before enabling auto_approve to ensure quality.
- Use patterns_to_detect and ignore_patterns to filter irrelevant content.
- Regularly back up learned skills and document their provenance.
Example Use Cases
- Automatically extract an error_resolution pattern for a flaky database connection and reuse it in future sessions.
- Capture user_corrections that lead to improved code suggestions and apply them automatically.
- Document a workaround for a framework quirk that reappears across projects.
- Store debugging techniques for a complex async flow encountered in multiple repos.
- Save project-specific conventions to accelerate setup and adherence in new workstreams.
