extract

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Extract Skill

提取Skill

Typically runs automatically via SessionStart hook.

Process pending learning extractions from previous sessions.

通常通过SessionStart钩子自动运行。

处理之前会话中待完成的经验提取任务。

How It Works

工作原理

The SessionStart hook runs:

bash

ao extract

This checks for queued extractions and outputs prompts for Claude to process.

SessionStart钩子会执行：

bash

ao extract

这会检查队列中的待提取任务，并输出提示信息供Claude处理。

Manual Execution

手动执行

Given

/extract

输入

/extract

后：

Step 1: Check for Pending Extractions

步骤1：检查待处理的提取任务

bash

ao extract 2>/dev/null

Or check the pending queue:

bash

cat .agents/ao/pending.jsonl 2>/dev/null | head -5

bash

ao extract 2>/dev/null

或者检查待处理队列：

bash

cat .agents/ao/pending.jsonl 2>/dev/null | head -5

Step 2: Process Each Pending Item

步骤2：处理每个待处理项

For each queued session:

Read the session summary
Extract actionable learnings
Write to
```
.agents/learnings/
```

对于队列中的每个会话：

读取会话摘要
提取可落地的经验总结
写入到
```
.agents/learnings/
```
目录

Step 3: Write Learnings

步骤3：写入经验总结

Write to:

.agents/learnings/YYYY-MM-DD-<session-id>.md

markdown

undefined

写入路径：

.agents/learnings/YYYY-MM-DD-<session-id>.md

markdown

undefined

Learning: <Short Title>

Learning: <简短标题>

What We Learned

我们的收获

<1-2 sentences describing the insight>

<1-2句话描述洞察内容>

Why It Matters

重要性

<1 sentence on impact/value>

<1句话说明影响/价值>

Source

来源

Session: <session-id>

undefined

Session: <session-id>

undefined

Step 3.5: Validate Learnings

步骤3.5：验证经验总结

After writing learning files, validate each has required fields:

Scan newly written files:

bash

ls -t .agents/learnings/YYYY-MM-DD-*.md 2>/dev/null | head -5

For each file, check required fields:
- Heading: File must start with
```
# Learning: <title>
```
  (non-empty title)
- Category: Must contain
```
**Category**: <value>
```
  where value is one of:
```
debugging
```
  ,
```
architecture
```
  ,
```
process
```
  ,
```
testing
```
  ,
```
security
```
- Confidence: Must contain
```
**Confidence**: <value>
```
  where value is one of:
```
high
```
  ,
```
medium
```
  ,
```
low
```
- Content: Must contain a
```
## What We Learned
```
  section with at least one non-empty line after the heading
Report validation results:
- For each valid learning: "✓ <filename>: valid"
- For each invalid learning: "⚠ <filename>: missing <field>" (list each missing field)
Do NOT delete or retry invalid learnings. Log the warning and proceed. Invalid learnings are still better than no learnings — the warning helps identify extraction quality issues over time.

写入经验总结文件后，验证每个文件是否包含必填字段：

扫描新写入的文件：

bash

ls -t .agents/learnings/YYYY-MM-DD-*.md 2>/dev/null | head -5

对每个文件，检查必填字段：
- 标题： 文件必须以
```
# Learning: <标题>
```
  开头（标题不能为空）
- 分类： 必须包含
```
**Category**: <值>
```
  ，其中值为以下之一：
```
debugging
```
  、
```
architecture
```
  、
```
process
```
  、
```
testing
```
  、
```
security
```
- 置信度： 必须包含
```
**Confidence**: <值>
```
  ，其中值为以下之一：
```
high
```
  、
```
medium
```
  、
```
low
```
- 内容： 必须包含
```
## 我们的收获
```
  章节，且标题后至少有一行非空内容
报告验证结果：
- 对于每个有效的经验总结："✓ <文件名>: 有效"
- 对于每个无效的经验总结："⚠ <文件名>: 缺失<字段>"（列出每个缺失的字段）
不要删除或重试无效的经验总结。 记录警告并继续执行。无效的经验总结总比没有好——警告有助于长期识别提取质量问题。

Step 4: Clear the Queue

步骤4：清空队列

bash

ao extract --clear 2>/dev/null

bash

ao extract --clear 2>/dev/null

Step 5: Report Completion

步骤5：报告完成情况

Tell the user:

Number of learnings extracted
Key insights
Location of learning files

告知用户：

提取的经验总结数量
关键洞察内容
经验总结文件的存储位置

The Knowledge Loop

知识循环

Session N ends:
  → ao forge --last-session --queue
  → Session queued in pending.jsonl

Session N+1 starts:
  → ao extract (this skill)
  → Claude processes the queue
  → Writes to .agents/learnings/
  → Validates required fields
  → Loop closed

会话N结束：
  → ao forge --last-session --queue
  → 会话被加入pending.jsonl队列

会话N+1开始：
  → ao extract（本Skill）
  → Claude处理队列
  → 写入.agents/learnings/目录
  → 验证必填字段
  → 循环完成

Key Rules

核心规则

Runs automatically - usually via hook
Process the queue - don't leave extractions pending
Be specific - actionable learnings, not vague observations
Close the loop - extraction completes the knowledge cycle

自动运行 - 通常通过钩子触发
处理队列 - 不要让提取任务处于待处理状态
具体明确 - 提取可落地的经验总结，而非模糊的观察
闭环管理 - 提取完成整个知识循环