
Six showcase posts first, then the "Fenjue" drop: how a skill that clones websites got 1,700 stars in a crowded niche
What it is An open-source skill for Claude Code and Codex. Give it a URL and it downloads the whole site, moves the code over line by line from the minified source into a new project, then uses five automated comparison checks (down to pixel-level) to prove the copy matches the original exactly.
Why it took off The author posted results of cloning flashy websites 6 times in linux.do, a Chinese developer community. He shared only the results, not the method, which built up a crowd waiting for "Fenjue"; open-sourcing it on the day paid that off. In an AI cloning space where everyone competes on "how close does it look", he won technical readers over with side-by-side diffs that prove correctness.
BHow to follow it Don't build another skill that clones a website. Take the playbook instead: pick something AI "doesn't quite get right", build a version you can prove with numbers (or a lightweight version of the heavy part it gets wrong), share the results in one community to build demand, then open-source it.
- Data (updated daily) · Repo created 2026-08-11 · 1,808 stars · 7 days +66
GitHub · stars Data 2026-10-09 → 2026-11-20 (43 days, 41 observations). Points before 10-28 come from deeply / archive backfill, not every day has one
website-rebuild-skill is an open-source skill that lets an AI agent clone a website line by line and then uses pixel-level comparison to prove the clone is correct. Its maker is a designer by background, with only this one repo. It picked up 1,736 stars and 312 forks in 93 days. Similar projects have stars in the tens of thousands, so it's an order of magnitude smaller. Still, its approach is worth unpacking: it posted a month of results in a Chinese developer community and only open-sourced once people asked for the method. In a space where everyone competes on "how close does it look", it competes on "can it prove it's right." Below: how it got going, what you can take from it, and whether it's worth following.
A post from midday, Aug 31
At noon Beijing time on Aug 2026, 31 12 linux.do user 80yan9 (nickname Boyang) posted in the development section, titled "[Open Source] The Fenjue for cloning websites is here!" The opening line of the body:
"I've shared a lot of website-cloning projects before, and got plenty of support and praise from fellow devs. Lots of you wanted the Fenjue, so I've now summarized and iterated it into a skill to share with everyone~"
He went on to explain why he built this. Over a decade ago he kept seeing great websites and wanted to learn how they were made, but "as a half-baked designer and 0.1-bit front-end cutter, I had no idea where to start." Asking the programmers around him felt like an imposition on their time. "The AI era changed that."
The first floor got 504 emoji reactions. The whole thread had 673 likes, 6,500 views and 114 replies (linux.do, 11-11). Most replies were short: "thanks, just what I was looking for", "starred to watch how it goes", "marking this". Someone in the thread said the repo had nearly 300 stars within 2 hours of the post. Three days later, the Chinese X account @GoSailGlobal reposted it with a side-by-side comparison video, by which point it had 706 stars.
"Fenjue" was built up over a month
The name wasn't picked at random. From Jul 20 to Aug 11,80, yan9 posted 6 threads on linux.do with diary-style titles: "I cloned a website, sharing it so we can learn", "Whenever I see a good site I want to clone it and learn", "Haven't cloned a site in days, here's another one", "At a fellow dev's request, another engineered clone of a great site"... Each got a few hundred to over a thousand views (the top one reached 1,295).
Those 6 posts only showed results, never method. He cloned award-winning sites with strong visuals: Lusion's 3D physics stacking, Samsy's cyberpunk WebGPU scene, F1 driver Lando Norris's personal site, and Shopify Editions. The comments kept asking how it was done, and "Fenjue" (a term for a martial arts secret manual) is the name that stuck.
So the star surge on Aug 31 wasn't caused by one post. It was "want" that built up over a month and finally paid out. It also explains why his V2EX post 0 got little reply that same day: nobody there was waiting for it.
What it actually is
Give an agent a URL and say "clone this site," and it walks through a fixed pipeline by itself: first it decides whether the site is feasible, then it captures the whole site as a read-only "evidence" set (every file gets a checksum), then it unpacks the minified code and moves it line by line into a new project, and finally five automated gates compare the two sides: console, network requests, DOM structure, element positions, and pixel-by-pixel screenshots. Each difference is either fixed or logged on record. "No glossing over."
A few rules are unusual. Every line in the clone must trace back to a specific line of the original code. The original's bugs and odd quirks are copied as-is, not fixed. Only in the final step is the verbatim-moved code reorganized into modules a human can read. The README's headline result: for a site with a custom WebGL engine, three pages match pixel for pixel, and the code is finally split into 389 source modules.
The toolchain is 75 zero-dependency Node scripts. One command, npx website-rebuild-skill, installs it, and both Claude Code and Codex can run it. The bar is not low: you need Node 22 and a local Chrome. Small to mid-size sites finish in one session. Heavy WebGL sites take 1 to 3 days.
The first screen of the README has not a single adjective. It shows about 6 side-by-side comparison images, original on the left and clone on the right, each labeled with meanAbsDiff 0.00 (the average per-pixel difference between the two images is zero). One of them is the Raycast keyboard page. The author points out a placeholder bar above the keyboard that never loaded, "identical on both sides," bug included. Further down is a section on "how this differs from scraping a site": a scraper saves what a site looks like, while a clone has to explain how it works. This is the point most questioned in the linux.do thread, and he wrote the answer into the docs himself.
Who it's competing with
"Have AI clone a website" is a crowded topic this 2026 year (GitHub stars, around 11-11):
| Project | star | How it differs |
|---|---|---|
| abi/screenshot-to-code | 81k | Screenshot-to-code, relying purely on visuals, never touching the original source |
| JCodesMore/ai-website-cloner-template | 39k, +405 in the last 7 days | One command extracts fonts, colors, and components and rebuilds them in React, aiming for speed and resemblance |
| firecrawl/open-lovable | 29k | Scrape the page, then have a model rewrite it. In essence, a redesign |
| Jakeschincariol/replica-skill | 9K, +7 over the last ~1,319 days | Launched in October, "clone any App and sell it," changing the pain points and rebranding |
| Jane-xiaoer/claude-skill-web-clone | 1,108 | Closest in philosophy ("real source code first"), but only a method doc, with no tooling or acceptance gates |
| HTTrack / SingleFile | 4.8K / 23k | Site archiving, saves only the look |
The leaders win on "one command, a few minutes, looks right," which means low barriers and a lot of people. This repo put itself in the narrowest and hardest slot: verifiable and traceable. That differentiation makes it memorable in a saturated field. The fork-to-star ratio reached 18%, which shows people are actually taking it and modifying it. But the same differentiation also means it is unlikely to reach tens of thousands of stars. The people willing to spend a day or two running a clone just to "learn how they did it" were never that many.
The "provably correct" idea is already being copied. albertirgi/replica was created on Nov 2 with the tagline "clone any UI and prove it with a score" and now has 106 stars. On Nov 9, someone posted on dev.to about a skill they built for pixel-perfect WebGL site cloning. In October 22, a post on cnblogs wrote up a skill along the same lines, with "three executable verification gates."
After launch: a slow long tail
The first three days brought in about 700 stars. The following two months ran on a long tail: dozens of skill directory sites auto-indexed it (it follows the Agent Skills spec, with both claude-code and codex in its tags), a batch of Chinese reposts and "quick start" articles in mid-October, and a GitCode mirror. GitHub's daily star counts show it went from 10-22 to 11-12 rising from 1,543 to 1,736, averaging 9 per day. The single-day peak was 10-18, with +55,11-08. There were also two small spikes, +11-12 and +36 on 38, whose sources I couldn't find. Over the last 7 days it gained +109, a bit faster than the +7 over the prior 61 days.
npx installer downloads (npm, 11-11) were 851 in September, 396 in October, and 11 in the first 102 days of November, so interest is fading. The last commit was on Oct 12.
It never made GitHub Trending, never made Hacker News, and got no real traction in the English-speaking world. Nearly all the growth came from the Chinese-speaking community. The author's only "monetization" is a promo code for an SSH client in the README's friendly links.
If it were you
One-line verdict: don't build this repo again. The "verifiable clone" slot is already taken, and others are already following it. What's worth taking is the playbook: serialize your results in a community to build demand, then change the evaluation criteria in a crowded topic. Anyone can use this approach, without big-company traffic or an existing following, but it takes over a month, and what you build has to be good enough to show. Worth trying.
Who it fits: developers or designers who already have a workflow that produces striking results (cloning, motion design, icon generation, shader writing all count) and who are willing to publicly show off their work for a month first, especially people who already have some presence in communities like linux.do, V2EX, or Jike.
Who it doesn't fit: anyone who wants to spend three days building "yet another AI website clone" to chase stars. The top of this topic is projects with tens of thousands of stars, and replica-skill, which launched in October, gained 1,300 stars in one week. Going head-to-head there, nobody will notice you.
How to do it:
- Use the same move in a different domain: pick something AI gets "looks wrong" at, and turn "does it look right" into something you can quantify and prove. Examples: design-to-code with pixel diffing, animation cloning with frame-by-frame comparison, email templates compared across client renderers. The linux.do post asking for help with "the AI output is way off from the real thing" applies to many topics.
- Do what it didn't do well. It's heavy, slow, and requires a local Chrome. A heavy site takes 1 to 3 days. Cloning just one section, one component, or one animation, getting results in minutes with a difference score, is far lower-barrier and reaches a much bigger audience (inference).
- Switch language communities: its growth came almost entirely from the Chinese-speaking community. Conversely, tools that are hot in the English world often don't yet have a clearly explained, verifiable version in the Chinese world.
- Borrow from its README: put the side-by-side comparison and a checkable number on the first screen, with no adjectives. Give "how this differs from XX" its own section, answering the most-questioned community objections up front. Install with a single npx command, follow the Agent Skills spec, and you'll get picked up by directory sites along the way.
How to get customers: serialize first, open-source later. In your target community, post "here's what I built now" every few days, showing only results. When readers ask about the method, say "I'll open-source it once it's cleaned up." On launch day, use a headline that pays off the promise to pull back readers from the earlier posts. His V2EX post got 0 replies, which shows that without the warm-up, a single post doesn't work. After launch, look for Chinese X accounts that repost GitHub projects and include a side-by-side comparison video.
When to do it, and when it's too late: the Agent Skills wave is still rising, directory sites are still auto-indexing new skills, and building now lets you catch that long tail. Once Claude Code and Codex ship built-in screenshot comparison and automated acceptance, "provably correct" won't be novel anymore.
Signs you've misjudged it: ① a "verifiable clone" project like albertirgi/replica that launched cold with no warm-up and still grows past a thousand stars shows demand is large enough that warm-up isn't the key, so you should just build directly; ② if this repo's daily gains drop to 0 to 2 without updates, its 1,700 stars were mostly a one-time spike on launch day, and the "verifiable" selling point doesn't keep people; ③ if leaders like JCodesMore or replica-skill add pixel-level acceptance, this slot's differentiation is gone.
Data and sources
GitHub repo page, README, commit history, and daily star counts (10-22 to 11-12); the linux.do launch post "[Open Source] The Cloning Manual Is Here!" (8-31, likes and views as of 11-11) and the author 80yan9's 6 cloning posts from July to August; V2EX "Share Creations" post 8-31; X @GoSailGlobal 9-03, @RepoGems 9-02; npm download counts (9-01 to 11-11); the author's homepage boyang.hu; competitor star counts as of 11-11 approximately, with "K" and "K" rounded.
Research notes: competitor table, acquisition evidence, full research report
- Competitors
Name Where Launched Scale How it differs from this one JCodesMore/ai-website-cloner-template Open source 2026-03-13 39,013 stars, +405 in the last 7 days One command extracts design tokens and rebuilds, aiming for speed and resemblance; no line-by-line traceability or pixel-level acceptance firecrawl/open-lovable Open source 2025-08-08 28,893 stars Scrape the page, then use an LLM to rewrite it as a modern React app. In essence, a redesign, not a restoration Jakeschincariol/replica-skill Open source 2026-10-03 8,983 stars, +1,319 in the last 7 days 11 skills that clone any App for commercial use (changing pain points, rebranding), not aiming for byte- or pixel-level fidelity abi/screenshot-to-code Open source 2023-11-14 81,160 stars Screenshot-to-code, purely visual inference, never touching the source site's code The web-clone skill in nexu-io/open-design Open source 2026-04 (repository) 107,240 stars on the repo A method prompt inside a large repo, with no toolchain or automated acceptance gates Jane-xiaoer/claude-skill-web-clone Open source 2026-05-28 1,108 stars Same philosophy (real source code first), but only a 6-step method doc, with no 75 scripts and no five-layer acceptance Scraping tools like HTTrack and SingleFile Open source Years ago 4.8k / 22.9k stars Only saves "what it looks like"; this project delivers readable module trees, reverse-engineering notes, and an acceptance report - Where users come from
- Verified linux.do serialized warm-up: between 2026-07-20 and 08-11, author 80yan9 posted 6 cloning-result threads in a row (426–1,295 views each), showing only results while readers kept asking for the "manual"
- Verified linux.do open-source launch post: 2026-08-31 "[Open Source] The Cloning Manual Is Here!" 673 likes, 6.5k views, 114 replies; a reply in the thread says "2 hours, nearly 300 stars"
- Inferred Trendshift daily chart: README badges #14 Repository Of The Day, #3 JavaScript (2026-08-31), the author put them in the README as social proof
- Verified Chinese X repost accounts: @RepoGems 2026-09-02, @GoSailGlobal 2026-09-03, with side-by-side comparison videos (the post says 706 stars)
- Verified Auto-indexed by skill directory sites: typingmind, skillsdirectory, skillsllm, bestofai, mcpmarket, skillpm, DeepWiki, thedailyrepo, and others
- Inferred Chinese SEO pages / GitCode mirrors: GitCode blog 10-11 two guides, rxzm.cn and gaystar.cn 10-13 and so on, timing close to 10-17, 10-18 single-day +55
- Verified V2EX: posted in the Share Creations section 2026-08-31, 0 replies, largely ineffective
boyang-hu/website-rebuild-skill: why it works
Research date 2026-11-11 / channel github / data card refreshed 2026-11-12
Bottom line first
This is a solo open-source repo that packages line-by-line website cloning as an Agent Skill. In 93 days since creation it reached 1,736 stars and 312 forks (GitHub API as of today: 1,738 / 313).
Why it took off: mostly a month of warm-up by the author on linux.do, plus a concentrated surge on launch day. The author Boyang (linux.do account 80yan9, a level-3 active user) comes from industrial design and product, a "semi-amateur designer" (boyang.hu/about describes him as Product Lead · Industrial Designer). Between 2026-07-20 and 08-11, he posted 6 "cloning result" threads in a row on linux.do's development and tuning board. They only showed off the flashy WebGL sites he cloned, with no methods included. Commenters kept asking how he did it, and he called the method the "manual." 08-31 He then packaged the method as a skill and posted "[Open Source] The Cloning Manual Is Here!" The opening post got 504 emoji reactions, and the whole thread had 673 likes, 6.5k views, and 114 replies. A reply in the thread said "nearly 300 stars in 2 hours." The same day it reached #14 overall on the Trendshift daily chart and #3 in JavaScript (per the README badge; the commit record says 2026-08-31). 09-02 and 09-03 the Chinese X accounts @RepoGems and @GoSailGlobal (Jason Zhu) reposted it, and the latter's post was at 706 stars. The next two months grew slowly through long-tail channels like skill aggregator sites, content farms, and GitCode mirrors, adding a dozen to a few dozen stars a day.
Differentiation: don't compete on "looks like", compete on "provably right". The rivals (JCodesMore/ai-website-cloner-template 39k stars, firecrawl/open-lovable 29k stars, Jakeschincariol/replica-skill 9k stars) all do the same thing: take screenshots, extract design tokens, and have AI rewrite a React site that looks similar. This repo flips it: it treats the source site's compressed bundle as the spec, ports it line by line, copies the bugs too, then runs five automated acceptance gates (console, network, DOM, geometry, pixel-by-pixel), and the first thing you see in the README is a side-by-side diff with meanAbsDiff 0.00. That difference is real and part of why it took off, but it isn't the only reason. It solves a concrete pain point in the Chinese dev community: "AI-cloned UIs look nothing like the original" (there are help threads about exactly this on linux.do). Add the trust the author had built beforehand, and you get the launch-day spike.
---
1. Data profile
| Metric | Value | Source |
|---|---|---|
| Repo created | 2026-08-11 | GitHub API |
| Public release | 2026-08-31 (linux.do / V2EX / NodeSeek same day) | linux.do topic 2834831;V2EX API t/1238422 |
| star / fork | 1,736 / 312(fork/star ≈ 18%) | Data card 2026-11-12 |
| last 7 days / previous 7 days | +109 / +61 | Data card |
| past 30 days | +336 | Data card |
| highest single day in our database | +55(2026-10-18) | Data card |
| our database 10-22 → 11-12 | 1,543 → 1,736, about 9 stars/day on average, with two small jumps: 11-08 +36 and 11-12 +38 | Data card daily |
| 3 days after launch | about 706 stars (figure from @GoSailGlobal 09-03 post) | x.com/GoSailGlobal/status/2095315870285439057 |
| 2 hours after launch | "near 300 stars" (a user in the thread) | linux.do 2834831, around floor 40 |
| npm installer downloads | 2026-09 851 / October 396 / November 1–11 days 102, total 1,349 | api.npmjs.org |
| Contributors | 3 (author + 2 community PRs, one of them Windows support) | GitHub API |
| last commit | 2026-10-12(v0.3.24) | commits |
| reached GitHub Trending | no (trending_days_any = 0) | Data card |
| HN | Never charted | hackernews database repo query returned nothing |
Note: GitHub currently returns 401/403 for the stargazers endpoint, so day-by-day star timing isn't available. The launch-period curve can only be pieced together from numbers in third-party posts: 2 hours about 300,3 days about 706, and the data card 10-22 is 1,543. From this, the first 3 days picked up roughly 40% of stars, and the 9–October long tail added about 45%.
2. Competitors
| Product | Form | Launched | Scale | How it differs from this project |
|---|---|---|---|---|
| abi/screenshot-to-code | Open source | 2023-11 | 81,160 stars | Screenshot-to-code, purely visual inference, never touching the source site's code |
| JCodesMore/ai-website-cloner-template | Open source (Claude Code template + /clone-website command) | 2026-03-13 | 39,013 stars, +405 in the last 7 days | One command to "clone any website": extracts fonts, colors, and components, then rebuilds them as Next/React with parallel agents. Optimized for speed and visual similarity, with no line-level traceability or pixel-level acceptance. A Zhihu article called it "surged 26k stars and hit Trending several times" |
| firecrawl/open-lovable | Open source (official Firecrawl) | 2025-08 | 28,893 stars | Scrapes a page, then uses an LLM to rewrite it as a modern React app. Essentially a redesign |
| Jakeschincariol/replica-skill | Open-source skill pack (11 skills) | 2026-10-03 | 8,983 stars, +1,319 in the last 7 days | "Clone any app", aimed at indie hackers building businesses: recon, rebuild, test, change user pain points, rebrand. A product-grade copy, not byte-level fidelity |
| The web-clone skill in nexu-io/open-design | A feature inside a larger repo | Repo 2026-04 | 107,240 stars on the repo | A "website cloning methodology" decision tree that tells the agent to find real source code or existing open-source templates. It's a prompt-level method, with no toolchain or acceptance gates |
| Jane-xiaoer/claude-skill-web-clone | Open-source skills | 2026-05-28 | 1,108 stars | A 6-step decision tree that insists on "real source first, never AI-invented code". Closest in philosophy, but only a methodology doc, with none of the 75 scripts or automated acceptance |
| ericshang98/Perfect-Web-Clone | Open source | 2026-01 | 283 stars | Paste a URL, get a measured Vite+React replica. Measured, but not ported line by line |
| oil-oil/draw-ui, KyrieCheungYep/ky-design-to-html-skill, and others | Open-source skills | 2026-04 / 06 | 553 / 165 stars | Screenshot-based UI recreation |
| HTTrack / SingleFile / wget | Open source / browser extensions | Years | 4.8k / 22.9k stars | Site archiving. The README has a dedicated section, "How is this different from site archiving": archiving answers "what does it look like", cloning answers "how was it built" |
| TIGER-AI-Lab/WebsiteBench | Academic benchmark | 2026-11-04 | 235 stars | Evaluates whether AI can rebuild a website through browsing, showing the academic community treats this as a serious problem |
Takeaway: the space is crowded, and the leaders (JCodesMore 39k, open-lovable 29k, replica-skill 9k) are each an order of magnitude bigger. It isn't the first or the largest. What it owns is the narrower, harder niche: forensic-grade, verifiable, line-level fidelity.
3. How users came in (growth levers)
| Channel | Evidence | Confirmed? |
|---|---|---|
| linux.do teaser posts (main driver) | The author 80yan9 posted 6 clone showcase posts between 2026-07-20 and 08-11: "Cloned a website, sharing for learning" (583 views), "Whenever I see a good site I want to clone it to study" (864), "Haven't cloned a site in days, here's another" (1,295), "Posting an engineered clone, from the community for the community" (426), "Per request, continuing to engineer-clone great sites" (925), "While testing the skill, here's another clone result" (638). Only showing results, building demand for "the secret manual". | Yes (linux.do topics/created-by/80yan9.json) |
| linux.do launch post (the trigger) | 2026-08-31 "[Open Source] The site-cloning secret manual is here!": a dev-tuned version, tagged "open source promotion", 673 likes, 6.5k views, 114 replies, 82 participants. A reply in the thread says "near 300 stars in 2 hours" | Yes |
| Cross-posted on V2EX / NodeSeek | V2EX "Share & Create" post 08-31, 0 replies. NodeSeek only appears as a friendly link in the README; the specific post wasn't found | V2EX confirmed invalid; NodeSeek not verified |
| Trendshift daily chart | README badges: #14 Repository Of The Day, #3 JavaScript (the commit message says 2026-08-31). The author put the badges in the README as social proof | The badges were added by the author; the Trendshift page wasn't read |
| Chinese X repost accounts | @RepoGems 2026-09-02; @GoSailGlobal (Jason Zhu) 2026-09-03, with a side-by-side video, saying "706 stars" | Yes (the post exists; engagement numbers weren't retrieved) |
| Skill aggregators / directories | typingmind skills, skillsdirectory, skillsllm, bestofai, agenticmart, mcpmarket, skillpm registry, DeepWiki, thedailyrepo (11-10), githubsignals, llmrepos, and other auto-indexers | Yes, all passive indexing |
| Chinese content farms and GitCode mirrors | Two GitCode blog posts 10-11 ("Complete Guide" and "Quick Start"), rxzm.cn 10-13, gaystar.cn 10-13, 80aj.com 08-31 reposts, a telegra.ph mirror post, mozhenzhen.com, skills.tenten.co (Traditional Chinese), hysenlabs 09-10 review | Yes. The batch of SEO articles from 10-11 to 10-13 lines up in time with the single-day +55 jump in 10-17 and 10-18, but causation is unverified. |
| npm distribution | 09-03 was listed in an npx one-click installer, with 851 downloads in September | Yes. This lowers the install barrier; it isn't an acquisition channel |
| HN / Reddit / Product Hunt / media | Nothing found | None |
| Paid acquisition / star farming | No evidence. The fork/star ratio of 18%, the real human replies in the thread, and genuine npm downloads don't look artificially inflated | — |
4. Why it took off
- The demand comes from one concrete pain point. linux.do already has plenty of help threads like "How do I perfectly clone a site's UI? What Antigravity / Kiro writes never looks anything like the original" (topic 1466507). AI-written frontends that don't look right are something Chinese vibe coders run into every day in 2026, and "pixel-perfect" hits that squarely.
- Show the result first, the method later. The author built a persona with a month-plus of 'clone results' posts: the works were award-winning WebGL sites like Lusion, Samsy, Lando Norris, and Shopify Editions, with strong visual impact. Readers kept asking for the 'secret method', so launch day became the moment of delivering on the promise, and the stars built up beforehand were released all at once.
- The README is evidence, not marketing. The first screen is about 6 side-by-side comparison images labeled meanAbsDiff 0.00; below that are 24 sites marked as fully cloned and 43 borderline samples graded as such, plus a full Codex run of Hashgraph VC (166/166 bytes identical). For technical readers, hard numbers they can check are far more credible than 'clone any website in seconds'.
- It rides the Agent Skills wave and works across runtimes. It follows the agentskills.io spec and claims to have been tested on both Claude Code and Codex. Its topics include claude-code / codex / agent-skills, so dozens of skill directory sites picked it up automatically, and some of the long-tail traffic comes from those sites.
- Bilingual and international distribution. It has an English README, and a Chinese X repost account picked it up. But no significant spread in the English community was visible; growth was mostly within the Chinese community.
5. Differentiation
- What it is: It treats cloning as 'forensics + reverse engineering + regression testing'. The full-site mirror is read-only evidence (a per-file sha256 ledger), compressed code is ported line by line, and each line maps back to the line number in the source site's bundle, with bugs copied rather than fixed. Five automated comparison gates either fix differences or log them. Only then comes 'source reconstruction', splitting the output into modules a human can read. It has 75 zero-dependency Node scripts and 23 scenario-specific guides, and it can rescue dead sites via the Wayback Machine and reconstruct flight streams for RSC (Next App Router) sites.
- Is that why it took off? Partly, but it determines whether it can be remembered in a crowded field, not how big it gets. Top competitors win with 'one command, a few minutes, looks similar', which has a low barrier and a broad audience. This repo's approach needs Node 22 and a local Chrome, and heavy WebGL sites take 1–3 days. Its audience is front-end developers and designers who want to learn how it was done, so the ceiling is clearly lower: 1.7k versus 39k. The differentiation gave it the spot of 'the hardcore one' in the Chinese community, and a fork/star ratio of 18% shows that people really take it and modify it.
- Others are already chasing the "verifiable clone" angle: albertirgi/replica (2026-11-02,106 stars, "Rebuild any UI from a reference and prove it with a score"); utkarshxdev/pixel-perfect-cloner (2026-11-09, dev.to post: "clones interactive WebGL sites pixel-for-pixel", 1 stars); a Cnblogs post, 2026-10-22 "Pixel-perfect website front-end replication," whose author built a skill on the same idea with "three executable verification gates." Caho1/open-claude-ui (1,745 stars) runs the same "reverse-engineer and restore + reverse-engineering playbook" story, but built on the REA tool.
6. How many people know about it
- It spread through a small circle of Chinese developers: a linux.do post with 6.5k views, two Chinese X repost accounts, plus a batch of SEO articles and aggregator sites. It did not reach HN, GitHub Trending, or English-language media, and no major influencer wrote an in-depth piece on it.
- No revenue involved (open source, MIT, no paid tier). The author's only monetization move was 2026-10-12 putting a promotional code for the Termark SSH client (affiliateCode) in the README's friendly-links section.
- The space itself is already widely discussed: JCodesMore appeared on Zhihu with '26k stars surge', and replica-skill gained 1,319 in the last 7 days. 'AI website cloning' as a broad topic is already crowded; the open slot is the 'provably accurate' layer.
7. What others can borrow
- Serialize the work to build demand first, then open-source the method. Post a few 'look what I built' threads in target communities (linux.do, V2EX and the like) and let readers ask for the method themselves. Conversion on launch day will be far higher than from a cold start.
- Use side-by-side comparison images and quantified metrics as the first screen of the README. 'Source on the left, clone on the right, diff 0.00' beats any adjective. Add a section on 'how this differs from XX (site scrapers)' to answer skeptics' questions up front.
- Change the evaluation dimension in a crowded field. Everyone competes on 'fast' and 'looks similar'; it competes on 'verifiable' and 'traceable', packaging engineering discipline (read-only mirror, difference log, acceptance gates) as the selling point.
- Install and use in one step: one npx command, sha256 verification after install, works across Claude Code / Codex, and gets auto-listed by skill directory sites.
8. What's uncertain
- Day-by-day star history isn't available (the stargazers endpoint returns 401/403); the launch-period curve can only be reconstructed from numbers in third-party posts.
- The specific sources of the 10-17, 10-18 (+55), and 11-08, 11-12 small jumps were not found. Guesses include GitCode and content-farm articles, skill directory featured slots, or a Chinese X account or WeChat Official Account I didn't capture. None of this is verified.
- The Trendshift daily ranking was inferred from the badge the author displayed; the Trendshift page itself was not read. The data card's claim of 'not in Trendshift top 25' doesn't match the badge (#14), which may be due to different scopes (all-language daily chart vs. long-term chart).
- Engagement and follower counts for the two X posts (@GoSailGlobal, @RepoGems) were not retrieved. The local x-radar corpus has no related posts.
- Real usage: npm downloads only count installs via npx; git clone and manual copies aren't counted. There are also no user-shared successful clone examples (most replies in the thread are 'starred, will try later').
- Carries copyright and compliance risk (cloning other people's websites). The author already made the defaults private, noindex, and non-deployed, and leaves legal judgment to users, but whether platforms will take it down is uncertain.
Sources
- GitHub repo and README: github.com/boyang-hu/website-rebuild-skill (README, CHANGELOG, commit history, contributors; read on 2026-11-11)
- Author homepage: www.boyang.hu/about
- linux.do launch post: linux.do/t/topic/2834831; author's post list: linux.do/topics/created-by/80yan9.json
- linux.do help thread (demand evidence): linux.do/t/topic/1466507
- V2EX: www.v2ex.com/t/1238422 (API shows 0 replies)
- X: x.com/GoSailGlobal/status/2095315870285439057 (2026-09-03), x.com/RepoGems/status/2094953121940136042 (2026-09-02)
- npm downloads: api.npmjs.org/downloads/range/2026-09-01:2026-11-11/website-…
- Competitor stars: our github repo_profile (2026-11-11) and the GitHub API
- Zhihu (JCodesMore competitor buzz): zhuanlan.zhihu.com/p/2058279875396493601
- Copycats: github.com/albertirgi/replica, dev.to/utkarshbhatnagar/i-built-an-ai-agent-skill-that-clone…, www.cnblogs.com/fhyy/p/23394641
- Long tail: blog.gitcode.com/4270d1788fb50efe34dadda9971786ba.html, hysenlabs.com/projects/boyang-hu-website-rebuild-skill, www.typingmind.com/skills/boyang-hu-website-rebuild
- Hacker News repo lookup:
hackernews repo boyang-hu/website-rebuild-skillreturned no results