Atlas / Skills / dylanma8232-art / daming-skillgain

daming-skillgainCAUTION

skills/dylanma8232-art/daming-skillgain

你的 Skill 到底有没有用?跑一次就知道——用真实任务做对照测试,看它在每个模型上是帮忙还是拖后腿。 | Measure whether a skill actually helps, per model, via paired with/without experiments and blind judge review.

Verdict
CAUTION
Grade
B
Trust score
82 /100
Version
1.0
Hosts
4 documented
License
GPL-3.0
Stars
0
01

Overview

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

English | 中文

写了几个 Skill,却说不清它们到底有没有帮上忙?同一个 Skill 换个模型可能从助力变成干扰;改完一版,也很难凭感觉判断结果究竟变好了还是变差了。

这是一个 Skill 发布前的质检门禁:把 Skill 放进真实任务里,与不启用时的产出对照;看完测评,再决定发布它,或在哪些模型上打开、关闭。它聚焦 Skill 的实际影响,不是通用 eval framework,也不负责替你优化 Skill 文本。

三个常见发布难题

  1. 第一次发布,不知道有没有用。 对同一模型运行相同任务,比较启用与不启用 Skill 的交付物,再依据规则检查和匿名评审决定是否值得发布。
  2. 换模型后,效果可能相反。 分模型展示同一任务的配对结果,找出哪些模型受益、哪些可能受损,避免把一个模型上的结果套用到所有模型。
  3. 修改 Skill 后,不知道变好还是变坏。 用相同任务和输入重新测修改版,查看两种条件下的产出、检查结果与评审证据,再决定是否采用新版本。

三种结论与测评流程

  • 有效:证据支持启用 Skill 后结果更好,可考虑在该模型上开启。
  • 有害:证据支持结果变差,可考虑在该模型上关闭。
  • 测不出来:有效证据不足或差异不明确;保留不确定性,补充测评再判断。
准备 Skill、真实任务和所需输入
│
选择要测的工具、模型与匿名评委
│
同一任务分别运行:不启用 Skill / 启用 Skill
│
执行确定性检查,评委匿名比较实际交付物
│
查看结论、并排产出和运行证据,再决定发布或开关

下载

直接下载可用 Skill 包(推荐给只想用、不改代码的人):

**⬇️ 下载 daming-skillgain v1.0**(纯 Skill 包:SKILL.md + references/ + evaluator/ + requirements.txt)

所有版本见 Releases。

unzip daming-skillgain-v1.0.zip
cd daming-skillgain
pip install -r requirements.txt
python3 -m evaluator.cli simple

30 秒上手(从源码)

需要 Python 3 和项目依赖;被测工具也需能在本机运行。也可以克隆仓库、安装依赖并启动交互式测评:

git clone https://github.com/dylanma8232-art/daming-skillgain.git
cd daming-skillgain
pip install -r requirements.txt
python3 -m evaluator.cli simple

按提示选择 Skill、任务、被测模型和评委,完成后打开生成的 report.html。

截图中的总结给出本次判断,按模型结果和并排产出展示差异与依据,帮助你决定是否应为该模型开启 Skill。

可信度与适用边界

  • 内置确定性检查按配置规则计算,不调用 LLM;检查规则和结果会保存为快照,支持复核。相同输入、规则与判定版本可重算出相同的规则判定。
  • 配对结论根据有效结果、差值、置信区间和证据门槛计算;证据未达到门槛时明确报告“测不出来”,不强行下结论。
  • 匿名评委只比较去掉条件标识的实际交付物;评委意见是判断证据之一,不等于客观真值。评委和模型生成结果可能变化,因此整体测评并非零 LLM。
  • 报告保留检查快照、评委原始意见、运行记录和产出文件,方便追溯本次判断。结果只适用于本次任务、模型、工具和输入条件。

支持范围与接入

Read from source at commit 2a0df2156adcOBSERVED · 2026-10-13
02

Install

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

pip install -r requirements.txt
git clone https://github.com/dylanma8232-art/daming-skillgain.git
pip install -r requirements.txt
pip install -r requirements.txt
git clone https://github.com/dylanma8232-art/daming-skillgain.git
pip install -r requirements.txt
03

Host compatibility

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

HostStatusNotes
claude-codementioned
codexmentioned
cursormentioned
windsurfmentioned
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: daming-skillgain
description: 用可复现的对照实验验证某个 Skill 是否真的有效——同模型 with/without 配对执行、确定性断言判定优先、评委盲评辅助,产出结果导向的静态 HTML 报告。当用户要判断 skill 有没有用、是否值得发布、换模型后是否失效、改版后变好还是变坏,或需要评测/回归某个 skill 时使用。
license: GPL-3.0
metadata:
  audience: developers
  workflow: evaluation
  version: "1.0"
  main_metric: Skill Gain Δ (paired, same-model)
  compatibility: 跨宿主(经 Subject Adapter 统一契约;报告标注本次已测宿主,结论仅适用于已测宿主)
---

# daming-skillgain

Skill 发布前的质检门禁:把 skill 放进真实任务,与不启用时的产出对照,再决定发布它、或在哪些模型上开启/关闭。

## 何时用我

- 判断某个 Skill 是否真的有效(而不是「看起来有效」)
- 找出 skill 在哪些模型上有正增益、哪些出现负迁移(应关闭)
- 在多个宿主上分别验证 skill 效果(结论不跨宿主外推)
- 对 skill 做版本回归:改版前后各跑一次同口径评测
- 发布 skill 前,需要一份可信、可复核的评测报告

## 最短路径(推荐)

```bash
pip install -r requirements.txt
python3 -m evaluator.cli simple
```

问答式引导,只答 4 个问题自动跑完评测,完成后打开生成的 `report.html`。

高级分步(`plan` → `run` → `evaluate` → `report`、参数细节、准备阶段硬规则)见 [references/workflow.md](references/workflow.md)。

## 关键约束(必须遵守)

- **评测需求必问**:先问用户「本次评测要让被测 skill 执行什么具体需求」。只有用户明确说「你根据 skill 自己生成需求」时才可省略 `--task-prompt`。
- **敏感资源不能生成**:账号、密码、token、API key、数据库连接、浏览器测试环境等只能由用户提供,不能由 AI 生成。
- **`needs_user_input` 必须停止**:计划生成返回 `needs_user_input` 时,把 `questions` 展示给用户并等待回答,不要继续 run。
- **证据不足不强判**:重复次数不足或置信区间跨 0 时,结论一律 `inconclusive`;`repetitions` 默认 3。
- **结论仅适用于已测宿主**:报告显式标注已测宿主清单,禁止跨宿主外推。

## 参考文件(按需读取,不要一次全读)

- [references/workflow.md](references/workflow.md) — 分步命令、参数、准备阶段硬规则、盲评与总结流程
- [references/usage.md](references/usage.md) — Python SDK / REST API / MCP Server 接入、宿主能力边界、执行器类型
- [references/hosts.md](references/hosts.md) — 接入被测 AI 工具的三种方式、工具矩阵、脚本契约模板
- [references/resources.md](references/resources.md) — 资源识别规则与预校验
- [references/report-and-contract.md](references/report-and-contract.md) — HTML 报告结构、评测契约与评测原则

## 前置条件

- Python 3.10+(推荐 3.12+)
- 至少一个能被程序唤起干活的 AI 工具:原生适配器宿主(如 OpenCode,开箱即用)、任何工具经脚本契约包装(见 hosts.md),或纯 API 宿主(`api`/`standalone` 需 provider key,不具备文件操作能力)
- 被测工具本身不需要装额外东西——评测器的 CLI 参数是评测器自己的入口
05

Trust audit

CAUTIONgrade B · trust 82/100 Install with care. The audit found things worth knowing before you trust its output.

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

What the source does

Filesystem
UNDECLARED (5 observation(s))
Network
declared (10 observation(s))
Shell
declared (2 observation(s))
Dependencies
not all pinned
Secrets in source
none-found

Findings (14)

MEDIUMFilesystem / path · fs.destructive · CWE-22, CWE-59
evaluator/engine.py:788
shutil.rmtree(path, ignore_errors=True)
MEDIUMFilesystem / path · fs.destructive · CWE-22, CWE-59
evaluator/engine.py:930
shutil.rmtree(p, ignore_errors=True)
MEDIUMFilesystem / path · fs.destructive · CWE-22, CWE-59
evaluator/engine.py:937
shutil.rmtree(p, ignore_errors=True)
MEDIUMFilesystem / path · fs.destructive · CWE-22, CWE-59
evaluator/engine.py:947
shutil.rmtree(workspace_skill_dir)
MEDIUMFilesystem / path · fs.destructive · CWE-22, CWE-59
evaluator/subject_adapter/__init__.py:130
shutil.rmtree(d, ignore_errors=True)
MEDIUMNetwork egress · net.raw_ip · CWE-200, CWE-319
evaluator/executors/opencode.py:35
def __init__(self, base_url: str = "http://127.0.0.1:3000", timeout: int = 300):
MEDIUMNetwork egress · net.raw_ip · CWE-200, CWE-319
evaluator/executors/opencode.py:101
url = f"http://127.0.0.1:{port}"
MEDIUMNetwork egress · net.raw_ip · CWE-200, CWE-319
evaluator/executors/opencode.py:110
return "http://127.0.0.1:3000"
MEDIUMNetwork egress · net.raw_ip · CWE-200, CWE-319
evaluator/executors/opencode.py:164
url = f"http://127.0.0.1:{port}"
MEDIUMNetwork egress · net.raw_ip · CWE-200, CWE-319
evaluator/executors/opencode.py:175
return f"http://127.0.0.1:{port}"
LOWInsecure crypto · crypto.weak_hash · CWE-327, CWE-338
evaluator/assertions/exact_match.py:224
hasher = hashlib.sha256() if algo == "sha256" else hashlib.md5()
LOWInsecure crypto · crypto.weak_hash · CWE-327, CWE-338
evaluator/reporting.py:393
anon_id = hashlib.md5(key.encode()).hexdigest()[:6].upper()
LOWInsecure crypto · crypto.weak_hash · CWE-327, CWE-338
evaluator/reporting.py:407
anon_map[key] = hashlib.md5(key.encode()).hexdigest()[:6].upper()
LOWSupply chain · supply.unpinned · CWE-829, CWE-1357
requirements.txt
pyyaml, openpyxl, requests, httpx
Why it matters. 4 requirement(s) not pinned with ==
Fix. pin exact versions

Gates applied: no_behavioural_pass.

Audited 2026-10-13 · audit v0.4.1 · source sha 2a0df2156adcfull audit observations/trust-audit/skill/dylanma8232-art__daming-skillgain.json · Report an issue / request a re-scan
06

Audit history

Every audit this skill has had.

DateSourceVerdictGradeScoreChange
2026-10-132a0df2156adcCAUTIONB82first audit
07

Questions

What does the daming-skillgain skill do?

你的 Skill 到底有没有用?跑一次就知道——用真实任务做对照测试,看它在每个模型上是帮忙还是拖后腿。 | Measure whether a skill actually helps, per model, via paired with/without experiments and blind judge review.

Is daming-skillgain safe to install?

With care. The audit graded it B (82/100) and found 14 things worth knowing before you trust this skill, listed below with the exact line each was found on.

What can daming-skillgain access on my machine?

The audit observed that it reaches the network, runs shell commands and reads or writes files. 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 daming-skillgain work with?

Its documentation mentions claude-code, codex, cursor and windsurf. 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 (2a0df2156adc), read on 2026-10-13. The repository is watched, and a new audit runs when it changes — this is the first audit.

Advertisement