技能优化器
- 作者仓库星标 366
- 作者仓库 awesome-legal-skills
- 领域
- 工程开发
- 作者声明的智能体
- 未找到明确声明;不据此推断已兼容或已测试
- 静态检查
- 88 / 100 · 启发式扫描,不代表运行安全
- 运行环境证据
- 未执行;当前没有隔离运行测试证据
- 作者 / 版本 / 许可证
- @lawve-ai · 未声明许可证
- 上下文用量评级
- 低消耗
- 接入复杂程度
- 即装即用
- 是否需要外部接口密钥
- 不需要
- 兼容的系统
- 未声明(默认跨平台)
- 运行环境要求
- 未检出明确要求
- 检测到的文件与系统行为
-
- 只读
- 允许写入 / 修改
- 网络行为
- 仅限本地
- 安装命令数
- 无(仅作为资料)
需要注意: 未声明工具权限范围;实际权限由运行环境决定。
第 1 步:确定技能
如果未提供技能名称,请列出`skills/`目录中的可用技能并询问:
第 2 步:检测信号
扫描对话中的信号 - 用户表达反馈的时刻: 信号类型 · 示例纠正 · “不”、“那不对”、“缺少 X”、“总是做 Y”,用户重写输出 成功 · “完美”、“是”、“完全正确”,用户接受而不进行更改 边缘情况 · 用户需要解决方法,技能无法处理请求
第 3 步:评估每个信号的质量
对于每个校正信号,评估它是否可以成为良好的技能指令。质量标准 1. 完整 该说明包含应用该说明所需的所有信息。无需寻找其他地方或做出假设。等级·示例
第 4 步:对信号进行分级
满足标准 · 行动 所有 4 个标准均通过 · 直接添加到技能 少于 4 个标准 · 要求澄清(参见步骤 5)
第 5 步:要求澄清
当反馈不符合所有标准时,请使用`AskUserQuestion`工具询问缺少的内容:如果用户提供了说明→更新说明并继续执行步骤6。如果用户更喜欢原始说明→按照原始说明继续执行步骤6。
第 7 步:如果获得批准
更新 SKILL.md 阅读`skills/[skill-name]/SKILL.md` 在相应部分添加每条指令 每条指令必须可读且适用于其自身 更新`skills/[skill-name]/CHANGELOG.md`
第 8 步:保存观察结果
对于无法处理的信号,提出保存:如果是,追加到`skills/[skill-name]/OBSERVATIONS.md`
# Self-Improve Skill
Analyze the current conversation and propose improvements to skills based on corrections, successes, and edge cases discovered during the work session.
## Triggers
- `self-improve` - Analyze session and propose improvements
- `self-improve [skill-name]` - Target a specific skill
- `self-improve on` - Enable automatic mode (hook)
- `self-improve off` - Disable automatic mode
- `self-improve status` - Show automatic mode status
- `self-improve [skill-name] history` - Show modification history
---
## Main Workflow (`self-improve`)
### Step 1: Identify the Skill
If skill name not provided, list available skills from `skills/` directory and ask:
```
Which skill should I analyze for this session?
[List skills found in skills/ directory]
```
### Step 2: Detect Signals
Scan the conversation for **signals** - moments where the user expressed feedback:
| Signal Type | Examples |
|-------------|----------|
| **Correction** | "No", "That's not right", "It's missing X", "Always do Y", user rewrites output |
| **Success** | "Perfect", "Yes", "Exactly", user accepts without changes |
| **Edge case** | User needed a workaround, skill couldn't handle the request |
### Step 3: Evaluate Each Signal for Quality
For each correction signal, evaluate if it can become a good skill instruction.
#### Quality Criteria
**1. COMPLETE**
The instruction includes all information needed to apply it. No need to look elsewhere or make assumptions.
| Grade | Example |
|-------|---------|
| Pass | "Structure output as: Key Terms / Risk Areas / Suggested Revisions" |
| Fail | "Use the standard format" (which format?) |
| Fail | "Follow our firm's guidelines" (what guidelines?) |
**2. PRECISE**
No vague or subjective terms. Two different people reading the instruction would understand it the same way.
| Grade | Example |
|-------|---------|
| Pass | "Flag non-compete clauses over 12 months as high risk" |
| Fail | "Be more thorough in the analysis" |
| Fail | "Make it more appropriate for clients" |
**3. ATOMIC**
One instruction addresses one single requirement. Multiple checks should be split into separate instructions.
| Grade | Example |
|-------|---------|
| Pass | "Check for governing law clause" |
| Fail | "Check for governing law, jurisdiction, and arbitration clauses" (three checks - split them) |
**4. STABLE**
If referencing regulations or standards, specify the version or date. The instruction should be evaluable the same way regardless of when it's read.
| Grade | Example |
|-------|---------|
| Pass | "Review the termination provisions under our internal policy [policy name and reference], dated December 12, 2024." |
| Fail | "Follow latest market standards" (which standards? will change over time) |
### Step 4: Grade the Signal
| Criteria Met | Action |
|--------------|--------|
| **All 4 criteria pass** | Add to skill directly |
| **Less than 4 criteria** | Ask for clarification (see Step 5) |
### Step 5: Ask for Clarification
When feedback doesn't meet all criteria, ask for what's missing using the `AskUserQuestion` tool:
```
I detected a correction but need more information to improve the skill.
You said: "[user's feedback]"
To create a clearer instruction, I need the following information:
[Structured tool call listing what's missing based on failed criteria]
```
**If the user provides clarification** → Update the instruction and proceed to Step 6.
**If the user prefers the original** → Proceed to Step 6 with the original instruction.
### Step 6: Propose Changes
```
--- Learning: [skill-name] ---
Proposed additions:
1. "[exact instruction to add]"
Source: "[quote from conversation]"
2. "[exact instruction to add]"
Source: "[quote from conversation]"
---
Apply these changes? [Y/n]
```
### Step 7: If Approved
1. **Update SKILL.md**
- Read `skills/[skill-name]/SKILL.md`
- Add each instruction in the appropriate section
- Each instruction must be readable and applicable on its own
2. **Update `skills/[skill-name]/CHANGELOG.md`**
- Create if doesn't exist
- Add new entry AT THE TOP:
```markdown
## [DATE (format: "January 7, 2026")]
[Description of changes in natural language, 1-3 sentences]
```
- Entry rules:
- Most recent at top
- 1-3 sentences max
- Natural language
- No git references
### Step 8: Save Observations
For signals that couldn't be processed, offer to save:
```
Save these observations for later review?
- "[signal 1]" - Status: [why insufficient]
- "[signal 2]" - Status: [why insufficient]
```
If yes, append to `skills/[skill-name]/OBSERVATIONS.md`
---
## Secondary Commands
### self-improve on
1. Run:
```bash
rm -f ./.disabled
```
2. Reply: "Automatic mode enabled."
### self-improve off
1. Run:
```bash
touch ./.disabled
```
2. Reply: "Automatic mode disabled."
### self-improve status
Check `.disabled` file existence and report.
## self-improve [skill-name] history
1. Display CHANGELOG.md content
2. Ask: "Would you like to revert to a previous version?"
3. If yes:
- update the appropriate sections in `skills/[skill-name]/SKILL.md`
- update `skills/[skill-name]/CHANGELOG.md` with a rollback note
---
## Examples
### Example 1: All criteria met
**User said:** "Always flag non-compete clauses over 12 months as high risk"
**Evaluation:**
- Complete: Yes - instruction is fully specified
- Precise: Yes - "12 months" and "high risk" are clear
- Atomic: Yes - single check
- Stable: Yes - no time dependency
**Result:** Add directly
### Example 2: Missing criteria
**User said:** "Flag any non-market-standard indemnification clause"
**Evaluation:**
- Complete: No - "non-market-standard" is not defined
- Precise: No - "market standard" is subjective and varies by deal type
- Atomic: Yes - single check
- Stable: No - market standards evolve over time
**Action:** Ask for clarification using the `AskUserQuestion` tool:
```
I detected a correction but need more details.
You said: "Flag any non-market-standard indemnification clause"
To make this actionable, can you specify:
- What makes an indemnification clause "non-market-standard"? (e.g., uncapped liability, coverage of indirect damages, no carve-outs for gross negligence)
Do you want to provide more details, or should I add the instruction as you stated it?
```
**If user clarifies:** Update the instruction and add it.
**If user prefers the original:** Add the instruction as stated.
---
## Important Notes
- Never guess what the user meant - always ask if unclear
- Never infer requirements from context - they must be explicit
- One instruction = one check - split bundled feedback
- Fewer good instructions is better than many vague ones
- CHANGELOG.md is the user-facing record
章节与文件索引
按说明文件中的顺序排列。
触发器
第 1 步:确定技能
第 2 步:检测信号
第 3 步:评估每个信号的质量
第 4 步:对信号进行分级
文件/命令 -> self-improve · self-improve [skill-name] · self-improve on · self-improve off · self-improve status · self-improve [skill-name] history · skills/ · AskUserQuestion
内容校验值 -> 4a50c2d494f8
当前原文的自动摘译,尚未人工审校。
触发器
self-improve- 分析会话并提出改进建议self-improve [skill-name]- 针对特定技能self-improve on- 启用自动模式(挂钩)self-improve off- 禁用自动模式self-improve status- 显示自动模式状态