rules-capture

规则捕获技能

契约

输入：

• 表达可复用规则、偏好、标准或更正的用户陈述
• 可选目标范围：用户、仓库、项目、技能或会话

输出：

• 已捕获的待定规则，包含脱敏后的支持性引述、提取出的规则、范围和状态
• 永久存储位置建议

创建/修改：

• 将待定捕获内容追加到 .agents/memory/captured-rules.md（标记为 status: temporary）
• 只有在明确确认后才添加永久规则
• 绝不逐字写入类似机密的值。保存或复述捕获内容前，先对令牌、密码、API 密钥、Cookie 和私密凭据进行脱敏。

外部副作用：

• 无

需要确认：

• 将待定规则提升为永久项目规则或用户规则之前
• 根据捕获的规则编辑共享/公开技能之前

委派给：

• 当规则是应转化为技能的可复用工作流时，委派给 skill-capture
• 当规则可能影响多个配置文件时，委派给 agent-config-audit

---

此技能的触发条件

当用户提到以下任一内容时，自动激活：

直接表达规则

• “始终执行 X”/“绝不执行 X”
• “我更喜欢 X”/“我不喜欢 X”
• “永远不要执行 X”/“停止执行 X”
• “从现在起执行 X”/“今后执行 X”
• “这不是我们的做法”/“我们就是这样做的”
• “规则是 X”/“标准是 X”

编码风格表达

• “使用 X 模式”/“不要使用 X 模式”
• “按 X 的方式命名”/“按 X 的方式格式化”
• “从 X 导入”/“不要从 X 导入”
• “约定是 X”
• “遵循 X 模式”

表达不满的迹象（尤其要捕获！）

• “你为什么要执行 X”/“停止执行 X”
• “我之前告诉过你”/“我已经说过了”
• “这错了”/“这不对”
• “修复这个”/“更改这个”
• 任何带有指令的脏话（强烈信号！）

关于标准的问题

• “X 的规则是什么”/“我应该如何执行 X”
• “应该是 X 还是 Y”/“哪种方式才正确”
• “X 的编码风格是什么”

---

捕获流程

检测到规则/偏好时：

1. 确认已检测到

I noticed you expressed a preference/rule. Let me capture this.

2. 提取规则详情

解析用户陈述，识别：

• 类别：适用于哪个领域？（编码、命名、导入、模式、工具、沟通）
• 规则类型：始终、绝不、偏好、避免
• 具体操作：具体应该/不应该做什么
• 上下文：适用的情形
• 示例：如果有提及，给出正例与反例

3. 记录到捕获文件

追加到 .agents/memory/captured-rules.md（如果文件不存在，则创建文件并添加包含 status: temporary 的 YAML 前置内容块）：

### [YYYY-MM-DD HH:MM] - [Category]: [Short Title]

**User said (redacted):**

> "[Short supporting quote with secrets replaced by [REDACTED_SECRET]]"

**Rule extracted:**

- **Type**: [ALWAYS | NEVER | PREFER | AVOID]
- **Action**: [What to do/not do]
- **Context**: [When this applies]
- **Category**: [coding | naming | imports | patterns | tools | communication | workflow]

**Example:**

```[language]
// Good
[good example]

// Bad
[bad example]
```

状态：待审核

### 4. Confirm with User

已捕获此规则：

• [规则简要摘要]

要将此规则添加到永久规则中吗？[是/否/修改]

### 5. On Confirmation
- **Personal preferences** → append to the user's platform-level instruction file under the appropriate section
- **Shared project coding standards** → append to the applicable `AGENTS.md`
- **Scoped project overrides** → use `AGENTS.override.md` only when inherited guidance must be replaced for a subtree
- **Claude-specific behavior** → append to `CLAUDE.md`
- **Durable project facts** (architecture, deployment notes, gotchas) → add/update a file in `.agents/memory/`
- Remove the entry from `.agents/memory/captured-rules.md` (or mark it PROCESSED)

---

## Escalate Before Documenting

A rule that repeats, or that an agent keeps breaking, belongs in structure, not prose. Before recording it as permanent text, check whether architecture, types, a lint that names the fix, or a behavior test can make the mistake impossible. Read [correct procedure](references/correct-procedure.md) for the ladder, the proof that each check fails on a real past mistake, and the rule-to-enforcer table. Record the rule as text only for judgment calls nothing can enforce.

---

## Rule Categories

### Coding Rules
- Pattern preferences
- Import conventions
- Naming conventions
- File structure
- Component patterns

### Workflow Rules
- Build/test preferences
- Git workflow
- Documentation requirements
- Communication style

### Tool Rules
- CLI preferences
- IDE settings
- Terminal behavior
- Background processes

### Communication Rules
- Response style
- Verbosity preferences
- Explanation depth
- Acknowledgment patterns

---

## Storage Locations

| Rule Type | Storage Location |
|-----------|------------------|
| Personal preferences | User-level platform instruction file |
| Project coding standards | Repo-level agent instruction file |
| Durable project facts | `.agents/memory/<topic>.md` |
| Pending review | `.agents/memory/captured-rules.md` (`status: temporary`) |

---

## Redaction Rules

Before writing or displaying a captured quote:

- Replace API keys, tokens, passwords, cookies, private URLs, and credentials
  with `[REDACTED_SECRET]`.
- Keep only the shortest quote needed to justify the rule. Summarize the rest.
- If the rule itself is about handling a secret, capture the policy, not the
  secret value.

## Auto-Capture File Format

File: `.agents/memory/captured-rules.md`

```markdown
---
status: temporary
last_verified: YYYY-MM-DD
---

# Captured Rules - Pending Review

Rules automatically captured from conversations. Review and promote to permanent storage.

---

## Pending Rules

[Auto-captured rules appear here]

---

## Processed Rules

[Rules that have been promoted to CLAUDE.md or .agents/memory/]
```

---

## Integration with Other Skills

This skill works with:

- **docs**: For formatting captured rules properly

---

## Cleanup Command

When user says "clean up rules" or "process captured rules":

1. Read `.agents/memory/captured-rules.md`
2. Present each pending rule for review
3. User can: Approve, Modify, Delete
4. Approved rules get promoted to CLAUDE.md or `.agents/memory/` as appropriate
5. File is cleaned up

---

## Example Interaction

**User**: "stop using console.log for debugging, use the logger service"

**Agent**:

```
I noticed a new rule. Let me capture it:

Rule: Use logger service instead of console.log for debugging
Category: Coding
Type: NEVER (console.log) / ALWAYS (logger service)

Added to captured rules. Should I promote this to permanent coding rules?
```

---

## Activation Triggers

The skill activates on these regex patterns:

```regex
# Direct commands
(always|never|don't|stop|from now on|going forward)\s+.+
(prefer|like|want|need)\s+(you to|it to|this to)\s+.+
(the rule is|the standard is|the convention is)\s+.+

# Frustration signals
(why (did|do) you|stop doing|I told you|I already said)\s+.+
(that's (wrong|not right|incorrect)|fix this|change this)\s+.+

# Questions about standards
(what's the (rule|standard|convention) for|how should I|which way is)\s+.+
```