Atlas / Skills / alchaincyf / 达尔文.skill 2.0

达尔文.skill 2.0BLOCK

skills/alchaincyf/darwin-skill

达尔文.skill —— 一个让你的Skill无限进化的系统:评估→改进→测试→保留或回滚 | Autoresearch-inspired autonomous skill optimization for Claude Code. Evaluate, improve, test, keep or revert.

Verdict
BLOCK
Grade
D
Trust score
68 /100
Version
—
Hosts
5 documented
License
MIT
Stars
6,464
01

Overview

From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.

English | 中文

动画由 huashu-design skill 制作

像训练模型一样优化你的 Agent Skills。

受 Andrej Karpathy 的 autoresearch 启发,将自主实验循环从模型训练搬到 Skill 优化领域。一个只能向前转的棘轮。

v2.0 · 更新于 2026-05-28 · 吸收微软研究院 SkillLens 与 SkillOpt 两篇论文做的系统性升级。

[](LICENSE) [](#whats-new-in-20) [](https://skills.sh) [](https://skills.sh) [](https://github.com/microsoft/SkillOpt)

npx skills add alchaincyf/darwin-skill
[!NOTE] 🤝 微软研究院把达尔文列进了 SkillOpt 的官方集成名单。 2026-06-03,微软在 SkillOpt 仓库 的更新里写道: 「gbrain, gbrain-evals, and darwin-skill have all integrated SkillOpt.」 我们吸收了它的 validation-gated 框架,它把达尔文写进了自己的集成名单。这是一次双向的致意。👉 去 SkillOpt 仓库看看

What's New in 2.0

2.0 不是缝缝补补,是系统性吸收微软研究院 2026-05-22 两篇论文后的结构性升级。五个变化:

1. 评分标准 8 维 → 9 维(吸收 SkillLens 实证的 73.8% rubric 药方)

  • 原「错误处理」维度升级为 失败模式编码 (Failure Mechanism Encoding):不只是「告诉 agent 别犯错」,而是把已知失败路径显式编码进 skill
  • 原「明确性」维度升级为 可执行具体性 (Actionable Specificity):明文禁止「建议/可以考虑/根据情况/灵活把握/视情况而定」等模糊词
  • 新增第九维 高风险行动黑名单 (High-Risk Action Blacklist):rm/git reset --hard/force push 等破坏性操作必须在 skill 中显式列禁

**2. 验证机制对齐 SkillOpt

Read from source at commit 42f1e05ebcd8OBSERVED · 2026-10-20
Advertisement
02

Install

Commands as the repository documents them. They are shown, not run.

npx skills add alchaincyf/darwin-skill
npx skills add alchaincyf/darwin-skill
npx skills add alchaincyf/darwin-skill
npx skills add alchaincyf/darwin-skill
03

Host compatibility

What the documentation claims. We have not run a compatibility test.

HostStatusNotes
claude-codementioned
codexmentioned
cursormentioned
gemini-climentioned
openclawmentioned
04

What it tells the agent

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: darwin-skill
description: "Darwin Skill 2.0 (达尔文.skill 2.0): autonomous skill optimizer, v2.0 integrates Microsoft Research SkillLens (arXiv 2605.23899) 9-dim rubric + SkillOpt (arXiv 2605.23904) validation-gated design + human-in-the-loop checkpoints. Evaluates SKILL.md files using a 9-dimension rubric (structure + effectiveness + meta-skill blacklists), runs hill-climbing with git version control, spawns independent judge agents for blind evaluation, validates improvements through test prompts with auto-break on diminishing returns, and generates visual result cards. Use when user mentions \"优化skill\", \"skill评分\", \"自动优化\", \"auto optimize\", \"skill质量检查\", \"达尔文\", \"darwin\", \"帮我改改skill\", \"skill怎么样\", \"提升skill质量\", \"skill review\", \"skill打分\"."
---

# Darwin Skill 2.0

> **v2.1 · 2026-06-10** — keep/revert 棘轮从「绝对分数 delta」改为「**paired 同-judge 比较 + 奇数 N 多数决**」(绝对分数 ±8 judge 噪音淹没保守编辑的真实增益、是 false-revert 源;within-judge 比较消除换尺污染)。绝对分数降级为 triage-only。
> **v2.0 · 2026-05-28** — 吸收 Microsoft Research SkillLens(arXiv 2605.23899)的 9 维评分药方 + SkillOpt(arXiv 2605.23904)的 validation-gated 验证机制 + human in the loop 三层守关。
>
> 借鉴 Karpathy autoresearch 的自主实验循环,对 skills 进行持续优化。
> 核心理念:**评估 → 改进 → 实测验证 → 人类确认 → 保留或回滚 → 生成成果卡片**
> GitHub: https://github.com/alchaincyf/darwin-skill

---

## 设计哲学

autoresearch 的精髓:
1. **单一可编辑资产** — 每次只改一个 SKILL.md
2. **双重评估** — 结构评分(静态分析)+ 效果验证(跑测试看输出)
3. **棘轮机制** — 只保留改进,自动回滚退步
4. **独立评分** — 评分用子agent,避免「自己改自己评」的偏差
5. **人在回路** — 每个skill优化完后暂停,用户确认再继续

与纯结构审查的区别:不只看 SKILL.md 写得规不规范,更看改完后**实际跑出来的效果是否更好**。

---

## 评估 Rubric(9维度,总分100)

> **设计依据**:基于 SkillLens 论文(arXiv 2605.23899)实证发现——LLM-as-judge 评估 skill 质量准确率仅 46.4%(接近随机),加入 meta-skill 三维度后提升到 73.8%。本 rubric 强化 dim3 / dim5 评分标准,新增 dim9「反例与黑名单」,权重平衡到 100。**目的:让评分对真实质量更敏感,减少 LLM judge 的乐观偏差。**

### 结构维度(59分)— 静态分析

| # | 维度 | 权重 | 评分标准 |
|---|------|------|---------|
| 1 | **Frontmatter质量** | 7 | name规范、description包含做什么+何时用+触发词、≤1024字符、**禁结尾加"灵活应用/根据情况判断"等空话尾巴** |
| 2 | **工作流清晰度** | 12 | 步骤明确可执行、有序号、每步有明确输入/输出 |
| 3 | **失败模式编码** | 12 | **必须显式编码失败模式**(写出"如果 X 失败 → Y"的明确分支);有fallback路径、错误恢复;**只写正向流程而不写失败分支扣 ≥3 分**(SkillLens meta-skill 维度) |
| 4 | **检查点设计** | 6 | 关键决策前有用户确认、防止自主失控;**检查点必须显性标记(🔴/STOP/CHECKPOINT),仅靠"如果...建议..."措辞不算** |
| 5 | **可执行具体性** | 18 | 不模糊、有具体参数/格式/示例、可直接执行;**禁止"建议/可以考虑/根据情况/灵活把握/视情况而定"等软化措辞**——出现 ≥3 处扣 ≥3 分(SkillLens actionable specificity 维度) |
| 6 | **资源整合度** | 4 | references/scripts/assets引用正确、路径可达 |

### 效果维度(35分)— 需要实测

| # | 维度 | 权重 | 评分标准 |
|---|------|------|---------|
| 7 | **整体架构** | 12 | 结构层次清晰、不冗余不遗漏、与花叔生态一致;**冗余/AI腔废话段落(说白了/换句话说/首先其次综上等花叔禁用词)出现一处扣 1 分** |
| 8 | **实测表现** | 23 | 用测试prompt跑一遍,输出质量是否符合skill宣称的能力 |

### Meta-skill 维度(6分)— 反例与黑名单

| # | 维度 | 权重 | 评分标准 |
|---|------|------|---------|
| 9 | **反例与黑名单** | 6 | **skill 必须有"不要做什么"的反例清单**;只写"应该做 X"没有"不要做 Y"扣 ≥3 分;红灯/危险动作/反模式应单独章节列出(SkillLens risk-action blacklist 维度) |

### 评分规则
- 维度1-7、9:每个维度打 1-10 分,乘以权重得到该维度得分
- 维度8(实测表现):跑2-3个测试prompt,按输出质量打1-10分
- **总分 = Σ(维度分 × 权重) / 10**,满分100
- ⚠️ **绝对总分只用于 triage(粗排「哪支最弱、先改谁」),绝不用于 keep/revert**。实测:同一份**未改**文字换个 judge 评,总分可摆 **±8**(一支只加了 3 个 🔴 字元的 skill、单评却 −8.5,全是 judge 换尺、非真实退步)。keep/revert 一律走 **Phase 2 的 paired 比较**。
  - **为什么**:LLM judge 给的是**抽样、不是测量**——分数住在「文字 × 该 judge 当下选的标准」里,不是文字属性。绝对总分 = 用**两台未校准磅秤**量节食前后,差值大半是磅秤差;paired = **同一台磅秤**量前后,误差相减抵销。pairwise preference >> absolute scoring 是 LLM judge 的已知结论(RLHF 用 pairwise 不用绝对分同因)。

### Rubric 的实证基础

rubric 设计依据来自 **SkillLens 论文(arXiv 2605.23899)** + **本机 controlled study**:

- SkillLens 发现 LLM-as-judge 准确率仅 46.4%(接近随机),加入 meta-skill 三维度后升到 73.8%
- 本机对 huashu-research 做 4 类 degradation → 5 个独立 judge 盲测一致 V1>V2,Δ 均值 +46.5(5/5 high confidence)

**结论**:rubric 能识别 gross degradation,但 fine-grained quality difference 仍不可信,**重要决策必须人审**。

→ 详细论文证据 + 5 judges 完整数据 + HL 实战案例数字见 [references/skilllens-evidence.md](references/skilllens-evidence.md)

### 关于「实测表现」维度

这是与纯结构评分最大的区别。评分方式:

1. 为每个skill设计2-3个**典型用户prompt**(不是边缘case,是最常见的使用场景)
2. 用子agent执行:一个带skill跑,一个不带skill跑(baseline)
3. 对比输出质量,从以下角度打分:
   - 输出是否完成了用户意图?
   - 相比不带skill的baseline,质量提升明显吗?
   - 有没有skill引入的负面影响(过度冗余、跑偏、格式奇怪)?

若子 agent 不可用(超时/资源限制),退化为「干跑验证」:读完 skill 后模拟一个典型 prompt 的执行思路,判断流程是否合理;必须在 results.tsv 标注 `dry_run`。**dry_run 比例 > 30% → 评估失效警告**(来自本机 controlled study:dim8 实测维度权重 23%,无 full_test 验证时分数不可信)。

---

## Runtime 适配性审查(gate 项,独立于 9 维度评分)

skill 应当能在 Claude Code / Codex / Cursor / OpenClaw / Hermes / Gemini CLI / OpenCode 等 50+ skills-compatible runtime 通用——否则其他 agent 解析时会被「在 Claude Code 里」「Claude Code skill」等措辞误判为「不是给我用的」直接拒装(实例:nuwa-skill 因此被 Marvis agent 拒绝)。

### Phase 1 基线评估时强制跑一次红灯扫描

```bash
grep -nE "(在 Claude Code|Claude Code skill|Claude Code 用户|Cursor only|Codex 中|^\[!\[Claude Code|~/\.claude/skills/[a-z]|/plugin install\b)" SKILL.md README.md 2>/dev/null
```

输出非空 = 红灯命中,**但须先读命中行上下文排除假阳性**(grep 命令本身/反例引用/讲解该规则的元陈述=假阳性,记 `runtime_scan=false_positive` 不改;判别表见 references/runtime-neutrality.md)→ 确认是真红灯(指令性用法)才强制把 Phase 2 第一轮定为 P0「runtime drift 修复」(写入 results.tsv 的 note 列 `runtime_warn=N`)。

### 例外(允许的「Claude Code 痕迹」)

frontmatter 触发词、花叔生态内部 skill 名引用、明确标注 runtime-specific 章节、commit message——这些正当出现,不算红灯。

→ 红灯/绿灯完整对照表 + 例外清单详细规则 + Phase 1/2/3 各阶段审查时机见 [references/runtime-neutrality.md](references/runtime-neutrality.md)

---

## 自主优化循环

### Phase 0: 初始化

```
1. 确认优化范围:
   - 全部skills → 扫描 .claude/skills/*/SKILL.md
   - 指定skills → 用户指定列表
2. 创建 git 分支:auto-optimize/YYYYMMDD-HHMM
3. 初始化 results.tsv(如不存在)
4. 读取现有 results.tsv 了解历史优化记录
```

### Phase 0.5: 测试Prompt设计

在评估之前,为每个skill设计测试prompt。这步很关键——没有测试prompt,「实测表现」维度就打不了分。

```
for each skill:
  1. 读取 SKILL.md,理解它做什么
  2. 设计2-3个测试prompt,覆盖:
     - 最典型的使用场景(happy path)
     - 一个稍复杂或有歧义的场景
  3. 保存到 skill目录/test-prompts.json:
     [
       {"id": 1, "prompt": "用户会说的话", "expected": "期望输出的简短描述"},
       {"id": 2, "prompt": "...", "expected": "..."}
     ]
```

展示所有测试prompt给用户,**确认后再进入评估**。测试prompt的质量决定了优化方向是否正确。

### Phase 1: 基线评估(Baseline)— triage 用途

> **本阶段绝对分数是 triage 排名(决定先改谁),不是 keep/revert 基准**。judge 对 gross 差异
05

Trust audit

BLOCKgrade D · trust 68/100 Do not install this without reading the findings. The audit found something that could harm you or your machine.

LayerWhat it checksResult
L0Provenance & inventoryPASS
L1Static analysis of the codeFAIL
L2Instruction surface (what it tells the agent)WARN
L3Class-specific surfacePASS
L4Behavioural (sandbox)SKIPPED

What the source does

Filesystem
none-observed
Network
none-observed
Shell
UNDECLARED (4 observation(s))
Dependencies
pinned
Secrets in source
none-found

Findings (6)

HIGHCode injection · code.os_system · CWE-78, CWE-94, CWE-95
scripts/screenshot.mjs:29
const groot = execSync('npm root -g', { encoding: 'utf8' }).trim();
Why it matters. a shell string executor
HIGHCode injection · code.os_system · CWE-78, CWE-94, CWE-95
scripts/screenshot.mjs:86
execSync(`start "" "${outputPath}"`);
Why it matters. a shell string executor
HIGHCode injection · code.os_system · CWE-78, CWE-94, CWE-95
scripts/screenshot.mjs:88
execSync(`open "${outputPath}"`);
Why it matters. a shell string executor
HIGHCode injection · code.os_system · CWE-78, CWE-94, CWE-95
scripts/screenshot.mjs:90
execSync(`xdg-open "${outputPath}"`);
Why it matters. a shell string executor
MEDIUMPrompt injection · scope.undeclared_shell · CWE-94, CWE-1427
<declared scope>
shell use found in code, not declared in the description
Why it matters. the description does not admit a capability the code has
Fix. declare shell use in the description, or remove it
INFOInventory / provenance · inv.oversize · CWE-1104
assets/hero.gif
assets/hero.gif
Why it matters. 2914686 bytes not read

Gates applied: no_behavioural_pass.

Audited 2026-10-20 · audit v0.4.1 · source sha 42f1e05ebcd8full audit observations/trust-audit/skill/alchaincyf__darwin-skill.json · Report an issue / request a re-scan
06

Audit history

Every audit this skill has had.

DateSourceVerdictGradeScoreChange
2026-10-2042f1e05ebcd8BLOCKD68first audit
07

Questions

What does the darwin-skill — alchaincyf/darwin-skill skill do?

达尔文.skill —— 一个让你的Skill无限进化的系统:评估→改进→测试→保留或回滚 | Autoresearch-inspired autonomous skill optimization for Claude Code. Evaluate, improve, test, keep or revert.

Is darwin-skill — alchaincyf/darwin-skill safe to install?

No — not without reading the findings first. The audit graded it D (68/100) and found 4 critical or high issues in the source. Each one is listed on this page with the file and line it is on.

What can darwin-skill — alchaincyf/darwin-skill access on my machine?

The audit observed that it runs shell commands. At least one of those is not mentioned in its own description, which is itself a finding. Secrets in the source: none found.

Which assistants does darwin-skill — alchaincyf/darwin-skill work with?

Its documentation mentions claude-code, codex, cursor, gemini-cli and openclaw. That is what the text claims, not a compatibility test we ran.

How current is this page?

The grade is for one exact copy of the source (42f1e05ebcd8), read on 2026-10-20. The repository is watched, and a new audit runs when it changes — this is the first audit.

Add this badge to your README

Shows this page's audit result and links back to it. It updates when the audit does.

ModelsAtlas audit badge

Markdown
[![ModelsAtlas audit: 达尔文.skill 2.0](https://modelsatlas.com/skills/alchaincyf/darwin-skill/badge.svg)](https://modelsatlas.com/skills/alchaincyf/darwin-skill/)

Not your listing, or something out of date? Claim or update your listing.

Advertisement