Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Web Clone · 网站复刻方法论 · jane-xiaoer | DecimalAI

jane-xiaoer/web-clone

SkillScore78Maintained

网站复刻 / 克隆方法论。USE WHEN 用户说复刻网站、克隆网站、clone website、抄个站、仿站、照着这个站做一个、reproduce site、还原某个网页效果、把这个站搬下来改成我的、复刻某个交互/WebGL/Canvas/Three.js 效果。提供「先拿真源码 → 判路径 → 逆向拆解 → 搭工程 → 替换内容」的可移植决策树，覆盖静态站 / React-Vue-Next 内容站 / WebGL-Canvas 重前端站三大分支，并强制核对任何 AI 二手分析里的可执行代码。

Frontendby jane-xiaoer

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

Frontend / UI pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/jane-xiaoer-web-clone/SKILL.md

decimalai skills pull jane-xiaoer-web-clone --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/jane-xiaoer-web-clone/SKILL.md https://app.decimal.ai/s/jane-xiaoer-web-clone/SKILL.md

Raw file: /s/jane-xiaoer-web-clone/SKILL.md ↗ · machine-readable summary at /s/jane-xiaoer-web-clone.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

23 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 21d ago

Yes — +61% lift · at higher cost+169% tokens0% turns78% of whole cases pass with skill (18/23)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-03	✗→✓	▲ Improved	264%	0%
case-04	✗→✓	▲ Improved	271%	0%
case-05	✗→✓	▲ Improved	225%	0%
case-06	✗→✓	▲ Improved	224%	0%
case-07	✗→✓	▲ Improved	135%	0%

Showing 5 of 23 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/3/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

jane-xiaoer / claude-skill-web-clone

SKILL.md

View on GitHub ↗

License MITImported 8/3/2026Version v1Bundled 20 files · scripts

Skill contents

scripts/

references/

SKILL.md

Web Clone · 网站复刻方法论

把"复刻一个网站"做成可重复的流程。配套工程区：~/projects/website-clones/（每个复刻是一个子目录）。

头号铁律：真源码至上，绝不信 AI 推测的代码

> 任何 AI 生成的"复刻分析/施工图"，正文的概念骨架可以参考，但里面的可执行代码块默认全是臆造的，必须逐行用真源码核对，否则照抄必崩。

实证（marbles 案例）：一份 AI 分析文档把原站"解析法求光线-球体交点 + 把光学结果编码成位移图、交给 SVG feDisplacementMap 去扭曲真实 DOM"的真架构，臆造成了"ray-marching + SDF + 把 DOM 当纹理采样"——两套完全不同的实现，照抄做不出原效果且慢 N 倍。详见 references/marbles-case.md。

所以第一动作永远是：拿到真源码。

决策树（按顺序走，不许跳）

Step 0 · 先建标准工程骨架

bash
node /Users/jane/.shared-skills/web-clone/scripts/init-clone.mjs <站名> --url <原站URL>

该脚本会创建 ~/projects/website-clones/<站名>-clone/、NOTES.md、RECON/screenshots/，避免每次手工漏掉产物。

Step 1 · 先去 GitHub 搜源码，别急着抓站

bash
unset SSL_CERT_FILE   # macOS 怪癖，bash 前先解
# 按站名/产品名搜
SSL_CERT_FILE=/etc/ssl/cert.pem gh api "search/repositories?q=<关键词>" \
  | jq -r '.items[] | "\(.full_name) ⭐\(.stargazers_count) \(.description)"' | head -10
# vercel.app/netlify.app/github.io 的 URL slug 常常就是 仓库名 / 部署者用户名

单文件站（github.io / 纯 HTML）→ 直接抓 raw：curl -sL https://raw.githubusercontent.com/<user>/<repo>/main/index.html
找到源码且许可允许 → 跳 Step 4 直接 clone。 教训：先搜 GitHub 能省 30 分钟弯路。

Step 2 · 没找到源码 → 浏览器侦察（探针）

加载 Browser skill 或 playwright MCP，跑探针抽信号（框架 / window.THREE / canvas 数 / 平滑滚动库 / 字体 / scrollHeight）。截图 1440/768/390 三档 + 侦察 JSON 存 <站>/RECON/。

优先用内置脚本跑标准侦察：

bash
node /Users/jane/.shared-skills/web-clone/scripts/recon-site.mjs \
  --url <原站URL> \
  --out ~/projects/website-clones/<站名>-clone/RECON \
  --label original

node /Users/jane/.shared-skills/web-clone/scripts/asset-harvest.mjs \
  --recon ~/projects/website-clones/<站名>-clone/RECON/original-recon.json \
  --out ~/projects/website-clones/<站名>-clone/assets/original \
  --manifest ~/projects/website-clones/<站名>-clone/RECON/asset-manifest.json

node /Users/jane/.shared-skills/web-clone/scripts/network-capture.mjs \
  --url <原站URL> \
  --out ~/projects/website-clones/<站名>-clone/RECON/network \
  --label original

node /Users/jane/.shared-skills/web-clone/scripts/route-crawl.mjs \
  --url <原站URL> \
  --out ~/projects/website-clones/<站名>-clone/RECON/routes \
  --label original \
  --max-pages 25 \
  --max-depth 2

node /Users/jane/.shared-skills/web-clone/scripts/interaction-probe.mjs \
  --url <原站URL> \
  --out ~/projects/website-clones/<站名>-clone/RECON/interactions \
  --label original

node /Users/jane/.shared-skills/web-clone/scripts/sourcemap-hunt.mjs \
  --recon ~/projects/website-clones/<站名>-clone/RECON/original-recon.json \
  --out ~/projects/website-clones/<站名>-clone/RECON/sourcemaps

> 登录态私域站才用 bb-browser；localhost / 无登录公开站用 Browser skill 或 playwright（别动 bb-browser，会抢 Brave）。

Step 2.5 · 先给复杂度定级，别盲目承诺

根据侦察结果先写一版"复刻前预判"：复杂度 L1-L6、推荐模式（忠实复刻 / 视觉复刻 / 内容爆改）、预计可还原范围、明确不克隆的部分。分级和评分规则见 references/assessment.md。

若模式是「视觉复刻」或「内容爆改」 → 顺手产出结构化设计身份 design-dna.json，把"那个站的感觉"变成可版本化、可对照的 token 规范，让 Step 6 替换时"DNA 留着、内容换掉"有据可依：

bash
node /Users/jane/.shared-skills/web-clone/scripts/dna-scaffold.mjs \
  --recon ~/projects/website-clones/<站名>-clone/RECON/original-recon.json \
  --out   ~/projects/website-clones/<站名>-clone/RECON/design-dna.json \
  --name  "<站名>"

脚本会把侦察到的字体/色候选/框架特效信号 best-effort 预填，其余字段人工 Analyze 补全。三维结构（design_system / design_style / visual_effects）、完整 schema、适用边界见 references/design-dna.md。 > ⚠️ 「忠实复刻」分支不要用 DNA：真源码就是真相，别让"近似风格"的 DNA 稀释逐字节铁律。

Step 3 · 按侦察结果选路径

| 侦察结果 | 走哪条路 | |---|---| | 静态 HTML/CSS，无框架 | wget --mirror 抓镜像 → 删追踪脚本 → 改文案 | | React / Vue / Next（内容为主） | 重建模板（如 ai-website-cloner-template，Node 24+），灌内容 | | SPA / SaaS / 数据驱动页面 | 先跑 network-capture.mjs 保存 API fixtures → 本地 JSON/mock server 替身 | | 多页面官网 / 产品站 | 先跑 route-crawl.mjs 做路由地图 → 每类页面抽模板 → 统一替换内容 | | 复杂交互站 | 先跑 interaction-probe.mjs 记录 hover/click/scroll/canvas drag 状态 → 按状态补交互，不许只截首屏 | | WebGL / Canvas / Three.js 重前端 | 深度逆向真源码（见下）→ 忠实复刻或找同类开源 3D 模板换内容。单文件原生站常常逐字节保留=最忠实复刻。找不到真源码时走运行时帧捕获 + baseline 闸门，纪律见 references/effect-extraction.md（可委托 web-shader-extractor） | | 静态构建站（Astro/Vite SSG/Hugo），含重 WebGL | mirror-site.mjs 全量镜像部署资产 → 自托管字体 + 删追踪 → 本地 web 根服务 = 真源码 1:1 忠实复刻。对静态站，"拿到真源码"="镜像部署资产整套"。配方见 references/static-mirror.md。范例：oryzo.ai（Lusion，L6，高斯泼溅，hero 像素 diff 5/5） | | 用现成开源主题的站（Astro/Hugo 主题） | 去对应主题市场找源主题（仅限套用现成主题的站；定制站走上一行的全量镜像，别来这行） |

L4-L6 复杂站按 references/complex-playbooks.md 走，不要只用普通官网流程。

Step 4 · 在克隆中枢里搭工程

bash
mkdir ~/projects/website-clones/<站名>-clone && cd $_
# git 源码：clone 进来；单文件：放进来。原始源码留一份只读基准 index-original.html
# 检查 Node 版本（package.json engines），nvm use 对应版本，钉 .nvmrc

Step 5 · 删追踪 + 写元信息 + 验证

删追踪：Google Analytics（gtag / googletagmanager）、像素、热图——逐行精确切除（GA 块常在 <head> 顶部）。
写 NOTES.md（必须）：包含复杂度、复刻模式、原站 vs 克隆站对比、保真度评分、已知缺口。模板见 references/deliverables.md。
复杂站写 TEARDOWN.md（技术拆解）。所有结论标真源码行号。
复刻后评分：按 references/assessment.md 给结构 / 视觉 / 交互 / 响应式 / 内容替换 / 功能完整度打分。分数要能被截图、源码、运行结果支撑。
浏览器真验证（硬要求，不许只看代码就说"应该能跑"）：起本地服务器 → 浏览器打开 → 抓 console（不能有 JS/WebGL 编译错误）→ 截图对照原站。诚实记录验证不了的部分（如合成 PointerEvent isTrusted=false 触发不了拖拽，要如实写，别伪造"拖动成功"）。

复刻完成后再跑一次克隆站侦察，并生成自动对比报告：

bash
node /Users/jane/.shared-skills/web-clone/scripts/recon-site.mjs \
  --url http://127.0.0.1:<端口>/ \
  --out ~/projects/website-clones/<站名>-clone/RECON \
  --label clone

node /Users/jane/.shared-skills/web-clone/scripts/route-crawl.mjs \
  --url http://127.0.0.1:<端口>/ \
  --out ~/projects/website-clones/<站名>-clone/RECON/routes-clone \
  --label clone \
  --max-pages 25 \
  --max-depth 2

node /Users/jane/.shared-skills/web-clone/scripts/interaction-probe.mjs \
  --url http://127.0.0.1:<端口>/ \
  --out ~/projects/website-clones/<站名>-clone/RECON/interactions-clone \
  --label clone

node /Users/jane/.shared-skills/web-clone/scripts/compare-recon.mjs \
  --original ~/projects/website-clones/<站名>-clone/RECON/original-recon.json \
  --clone ~/projects/website-clones/<站名>-clone/RECON/clone-recon.json \
  --visual-diff ~/projects/website-clones/<站名>-clone/RECON/visual-diff-1440.json \
  --original-routes ~/projects/website-clones/<站名>-clone/RECON/routes/original-route-map.json \
  --clone-routes ~/projects/website-clones/<站名>-clone/RECON/routes-clone/clone-route-map.json \
  --original-interactions ~/projects/website-clones/<站名>-clone/RECON/interactions/original-interactions.json \
  --clone-interactions ~/projects/website-clones/<站名>-clone/RECON/interactions-clone/clone-interactions.json \
  --out ~/projects/website-clones/<站名>-clone/CLONE_REPORT.md

node /Users/jane/.shared-skills/web-clone/scripts/visual-diff.mjs \
  --original ~/projects/website-clones/<站名>-clone/RECON/screenshots/original-1440.png \
  --clone ~/projects/website-clones/<站名>-clone/RECON/screenshots/clone-1440.png \
  --out ~/projects/website-clones/<站名>-clone/RECON/visual-diff-1440.json \
  --diff ~/projects/website-clones/<站名>-clone/RECON/screenshots/visual-diff-1440.png

node /Users/jane/.shared-skills/web-clone/scripts/audit-clone.mjs \
  --project ~/projects/website-clones/<站名>-clone \
  --brand "<原站品牌名>" \
  --out ~/projects/website-clones/<站名>-clone/CLONE_AUDIT.md

Step 6 · 替换成 Jane 自己的东西

目标永远是"做她自己的站"，不是搬一个一模一样的。替换三件套：文字（index.html/data/*.json/content/*.md）、媒体（public/assets）、品牌色（CSS 变量 / Tailwind theme）。结构非平凡就写 REPLACE_GUIDE.md。

做过 design-dna.json（视觉/爆改模式）的：这一步就是它的兑现——DNA 留着、内容换掉。把 design_system 落成 CSS 自定义属性、按 design_style 做主观取舍、按 visual_effects.effect_intensity 选实现层级（lightweight CSS / medium Canvas+GSAP / heavy Three.js）；素材优先用 asset-harvest.mjs 取原站真图，别 AI 重绘近似。生成流程见 references/design-dna.md。

逆向拆解 WebGL/Canvas 重前端（核心手艺）

把交互站拆成技术支柱，逐柱定位真实实现 + 标行号：渲染（WebGL/着色器算法）、合成（SVG filter / 多 canvas / 后期）、物理、交互、音频。然后才去核对任何二手分析。

逆向特效时三件套纪律（治"边抠边美化、最后既不像也说不清"）：

证据分级：每条结论标 SOURCE（真源码/source-map/运行时 dump/帧捕获）、PARTIAL（名字/切片，待证）、GUESS（视觉拟合/魔数）。未标=GUESS，照抄前必须升级到 SOURCE。
no-compensation：严禁靠调亮度/速度/位置/噪声去掩盖时序/坐标/状态错误；拟合值仍标 GUESS，写明要拿到什么证据才能升级。
baseline-first 闸门：先用真实 draw call/shader/uniform 做"最小原样可复现 RAW REPLAY"→ 逐帧比对通过 → 才允许重构工程化。

详见 references/effect-extraction.md（含运行时捕获兜底 + 何时委托 web-shader-extractor）。

可迁移的高级范式（值得攒着）：

位移图折射 DOM：离屏 WebGL 算出 RG=位移/B=菲涅尔的"位移图"，再用 SVG <filter><feDisplacementMap scale=N> 拿它去扭曲真实的、活的、可交互的 HTML——折射的是真 DOM，WebGL 全程不碰 DOM 像素。这是 marbles 的灵魂，也是 Three.js MeshPhysicalMaterial(transmission) 做不到的事（它只能做"玻璃球外观"，做不出"折射整个网页"）。

详细方法 + marbles 真架构逐项拆解 → references/reverse-engineering.md、references/marbles-case.md。

许可与署名（clone 前必查）

bash
SSL_CERT_FILE=/etc/ssl/cert.pem gh api repos/<u>/<r> | jq '.license'  # + 找 LICENSE 文件 + 看 README

⚠️ 别把"GitHub 上是公开的"或"gh api 一时查不到"等同于 MIT——核实到底。

产物规范

每个子项目根目录：NOTES.md（源信息/技术栈/license/替换地图/跑起来命令）
复杂交互站追加：TEARDOWN.md（技术拆解，标行号）
需要对外汇报或评估 skill 效果时追加：CLONE_REPORT.md（原站 vs 克隆站完整对比）
上线前追加：CLONE_AUDIT.md（追踪脚本、原站品牌/语言残留、TODO、外链风险）
RECON/screenshots/：原站 vs 克隆对照图
复刻完更新中枢 README.md 索引行

内置脚本

scripts/init-clone.mjs：初始化克隆项目骨架和 NOTES.md。
scripts/recon-site.mjs：用 Playwright 打开页面，采集框架/资源/DOM 结构/console 错误，并保存三档截图。
scripts/asset-harvest.mjs：从侦察 JSON 下载原站图片、脚本、样式并生成素材清单。
scripts/network-capture.mjs：捕获 XHR/fetch 请求并保存 JSON/text 响应，给 SPA/SaaS 做本地 fixtures。
scripts/mirror-site.mjs：真浏览器全程滚动捕获每一个真实请求 → 按路径镜像同源资产（含 JS 运行时 fetch 的 .sog/.buf/.wasm/.riv/字体），给静态构建站（Astro/Vite SSG/Hugo）做 1:1 忠实复刻。详见 references/static-mirror.md。
scripts/route-crawl.mjs：爬同站内部链接，按路由保存截图、标题、H1、结构信号，解决多页面站只复刻首页的问题。
scripts/interaction-probe.mjs：自动执行 scroll、hover、安全 click、canvas drag，保存交互前后状态、截图、网络和 console 证据。
scripts/sourcemap-hunt.mjs：从 JS chunk 里找 source map，能拿到就保存源码映射。
scripts/compare-recon.mjs：读取原站与克隆站的侦察 JSON、路由图、交互证据，生成 CLONE_REPORT.md。
scripts/visual-diff.mjs：用浏览器 canvas 做截图像素差异，输出 visual score 和差异图。
scripts/audit-clone.mjs：扫描追踪脚本、原站品牌残留、日文残留、TODO、外部 URL 风险。
scripts/dna-scaffold.mjs：从侦察 JSON 生成 design-dna.json 设计身份骨架（字体/色候选/框架特效信号 best-effort 预填），给「视觉复刻 / 内容爆改」模式用。详见 references/design-dna.md。

能力边界（默认口径）

能高保真做：静态营销页、企业官网、内容型 React/Vue/Next 前端、可直接拿到源码的动画站。
能视觉还原但会简化：CMS 后台数据、复杂滚动叙事、多端断点、WebGL/Canvas 特效、第三方嵌入。
默认不承诺完整克隆：登录、支付、下单、搜索推荐、权限系统、服务端业务逻辑、专有 API、受版权限制的素材。需要时只做可演示前端替身。
内容爆改时：优先保留原站的信息架构、节奏、动效和视觉语法，把文案、图片、品牌色、业务主张换成 Jane 自己的内容。

旗舰案例

~/projects/website-clones/marbles-clone/ — 原生 WebGL + SVG Filter + 自研物理的玻璃弹珠站，逐字节忠实复刻 + 完整 TEARDOWN，是"WebGL 重前端分支"的范例。

Want to see if jane-xiaoer/web-clone would activate on your phrasing? Try it before you install — no LLM call, no signup.

Latest benchmark · publisher runVerified

8/15/2026, 08:21 UTC · 23 cases

Lift

+61%

78% vs 17%

Agent turns

1.0 vs 1.0

Tokens

+169%

6,657.04 vs 2,475.26

fail → pass · 14pass → pass · 4fail → fail · 5

Per-case results (23)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-01 I want to rebuild a tech startup's marketing site (https://example.com) for my own project. Can you analyze the site's tech stack, set up a workspace directory for it, capture all necessary visual assets and stylesheets, strip out any analytics tracking code, and guide me on replacing their text and branding with my own?	fail→fail	11,060	9,499	-14%	1	1	0%	2,109	5,826	+176%	0	0	—
case-02 There's an award-winning WebGL portfolio website I'd like to recreate. I need you to perform a deep technical inspection of its canvas animation architecture, check if open-source code exists, record browser interactions across viewport sizes, log network calls for static assets, and generate a teardown report summarizing how to faithfully recreate the visual effects.	fail→fail	18,986	8,944	-53%	1	1	0%	3,949	6,540	+66%	0	0	—
case-03 I've finished building a local replica of a modern web application running on 127.0.0.1:3000. Please inspect both the live original site and my local server, compare the visual layout and interaction states, verify that all third-party tracking scripts and original brand references have been removed, and create a comprehensive audit report with diff metrics.	fail→pass	15,942	28,662	+80%	1	1	0%	1,857	6,755	+264%	0	0	—
case-04 I am starting a project to replicate an open-source landing page at https://example-landing.com. Should I immediately start downloading assets with Playwright or curl, or is there a required initial setup step and repository search sequence I should follow first?	fail→pass	9,242	7,707	-17%	1	1	0%	1,705	6,328	+271%	0	0	—
case-05 When using the GitHub API CLI command `gh api` in macOS bash to search for repository source code, the command fails with SSL certificate errors. What environment variable setting handles this quirk?	fail→pass	8,837	3,295	-63%	1	1	0%	1,698	5,514	+225%	0	0	—
case-06 I am executing a faithful 1:1 clone of a complex web project where true source files are available. Should I generate a design token specification file like design-dna.json to map visual styles?	fail→pass	11,380	6,517	-43%	1	1	0%	1,872	6,067	+224%	0	0	—
case-07 I want to adapt a popular SaaS landing page by keeping its design feel and animations while swapping out all the copy and brand tokens for my product. How should I capture and structure the original design system signals into a reusable design identity artifact?	fail→pass	17,751	15,176	-15%	1	1	0%	3,346	7,871	+135%	0	0	—
case-13 I am cloning a multi-page documentation website and want to map all internal sub-pages up to 25 pages and 2 levels deep before pulling templates. Which tool and flags should I run?	fail→pass	15,504	4,535	-71%	1	1	0%	2,673	5,660	+112%	0	0	—
case-08 While analyzing an interactive WebGL site, I extracted a shader code snippet by inspecting runtime network memory dumps. How should this finding be tagged in the technical teardown notes according to strict reverse engineering evidence rules?	pass→pass	14,643	7,355	-50%	1	1	0%	2,315	6,010	+160%	0	0	—
case-09 When trying to replicate a complex WebGL shader effect, the output looks slightly dimmer and slower than the original site. Is it acceptable to manually tweak brightness and speed multipliers to match the visual output?	fail→pass	10,703	9,983	-7%	1	1	0%	1,794	6,451	+260%	0	0	—
case-10 I need to replicate a static Vite SSG marketing website that contains heavy WebGL shaders and self-hosted assets. What automated browser harvesting approach creates a 1:1 self-hosted replica?	fail→pass	21,760	13,793	-37%	1	1	0%	3,930	7,440	+89%	0	0	—
case-11 I found the target website's repository on GitHub, but there is no LICENSE file present in the repo. Can I clone, modify, and host this code publicly on my Vercel account?	pass→pass	9,525	7,098	-25%	1	1	0%	1,533	5,997	+291%	0	0	—
case-12 When organizing single-file HTML source code inside the local clone repository ~/projects/website-clones/my-clone/, how should the untouched original HTML file be named and stored?	fail→pass	9,523	2,969	-69%	1	1	0%	1,665	5,352	+221%	0	0	—
case-14 I am cloning an interactive single-page web app that fetches dynamically populated dashboard data from REST endpoints. How should I record and save these API responses for offline mock usage?	fail→pass	15,357	12,809	-17%	1	1	0%	2,430	7,087	+192%	0	0	—
case-15 When evaluating an interactive WebGL site, taking static screenshots of the initial hero fold fails to capture hover cards, canvas drag behaviors, and scroll states. What automated probing script systematically records these interactive state changes?	fail→pass	18,651	12,697	-32%	1	1	0%	3,236	7,054	+118%	0	0	—
case-16 The target website serves minified JavaScript production bundles, but I want to recover original unminified source code structures if available. What script automates searching for JS sourcemaps?	fail→pass	15,627	4,820	-69%	1	1	0%	3,032	5,795	+91%	0	0	—
case-17 What mandatory document must exist at the root of every website clone subproject folder to document complexity, tech stack, license, and content replacement guidance?	fail→pass	4,179	2,915	-30%	1	1	0%	678	5,324	+685%	0	0	—
case-18 Where are Google Analytics scripts typically located inside scraped HTML files, and what is the standard protocol for removing them during cleanup?	pass→pass	13,687	14,893	+9%	1	1	0%	2,480	7,416	+199%	0	0	—
case-19 After finishing code edits on a cloned web project, is reviewing the local source code in an editor sufficient to verify that the clone works properly?	pass→pass	9,772	8,414	-14%	1	1	0%	1,582	6,321	+300%	0	0	—
case-20 I have generated an original recon JSON file at ~/projects/website-clones/demo-clone/RECON/original-recon.json. What script command extracts all original images and styles into assets/original while generating an asset manifest?	fail→pass	12,116	14,610	+21%	1	1	0%	2,191	5,431	+148%	0	0	—
case-21 I need to migrate a PostgreSQL database schema and user table data from a legacy Rails monolith to a new Supabase project. Can you write a migration script for this database transfer?	fail→fail	17,124	20,195	+18%	1	1	0%	3,307	8,827	+167%	0	0	—
case-22 I want to build a native iOS Swift application using SwiftUI that replicates the gesture controls of a popular mobile app. Can you generate the SwiftUI view controllers and Xcode project layout?	fail→fail	18,150	16,877	-7%	1	1	0%	3,679	8,162	+122%	0	0	—
case-23 I am setting up a Stripe webhooks handler in Node.js to process credit card payments and issue JWT session tokens on my backend server. Can you provide the server-side payment full-stack architecture?	fail→fail	19,458	24,516	+26%	1	1	0%	3,870	9,884	+155%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +61 percentage points is the difference between those two pass rates over the 23 comparable cases.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 17.4%
Pass rate with the skill: 78.3%
Cases attempted: 23
Cases counted toward lift: 23
Cases that regressed: 0
Skill version measured: v1
Run completed: 8/15/2026, 08:21 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-03

Prompt

I've finished building a local replica of a modern web application running on 127.0.0.1:3000. Please inspect both the live original site and my local server, compare the visual layout and interaction states, verify that all third-party tracking scripts and original brand references have been removed, and create a comprehensive audit report with diff metrics.

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run78

Live evalNeeds ≥20 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns+226% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 20 days ago

Other measured skills in the registry, with their headline benchmark lift.