boyang-hu/website-rebuild-skill

This page was translated automatically. Technical details may differ from the original. View the original

1.6K stars Quality 73 Trend 43MIT

Replicate website Agent Skills: capture read-only images, restore compressed code line by line, and automatically compare and accept it. An agent skill that mirrors a website, rebuilds it from the minified code, and verifies the result with automated diffs.

Overview

It's not about "making a page that looks very similar." It treats the source site's implementation as a specification: first, it treats the entire site as read-only evidence, then reconstructs the logic line by line from the compressed code, and finally uses a whole set of automatic acceptance gates to prove that "the replication side and the source station are doing the same thing"—including pixel-by-pixel comparison. Designed according to the Agent Skills open specification, designed for **. Cross-runtime is not just a slogan but a test: the validation checklist includes both the goals completed by Claude Code and the goals of ** (Hashgraph VC, 166/166 response bytes consistent)—the same skill directory can run a complete pipeline in any runtime. - ——Before starting work, first check which category the target belongs to and whether it can be done; If it can't be done, directly explain the reason; it won't force the process to generate junk -——Capture the entire site as a read-only snapshot (file-by-file sha256 ledger, reference closure verification, authenticity check), can run even if offline -——Every line in the replica can point to which line in the origin station bundle; Bugs and strange copied styles without fixing - — Five-layer automatic comparison between console / network / DOM / geometry / ** — Differences either fixed or registered, no blurred — Rewritten by word-by-word ported to a readable project (remove modules, name by evidence, add source tracing notes), - — 75 Node scripts (51 process scripts and acceptance gates + 14 shared libraries + 10 source/reverse refactor), the entire pipeline before source code is not loaded with any npm packages - — Wayback archive rescue: anchor point + time window select a continuous moment, original byte completed as standard image, permanent hole truthfully registered; Five dead station rescue tests (four revived, one completed L3; A single image layer loses all confirmations, and failure forms are recorded) — Server-side component sources are not issued (React Server Components / Next.js App Router), but their full output (flight stream) is inlaid into each HTML page as a specification: from this, a constructable Next project is reconstructed and closed with a flight semantic gate. In actual tests, a Next 16/Turbopack blog site has identical route semantics for 18/18, and a 144 route reserver PASS 144/144; Blind reverse grading of open source code, structure ≈95%, behavior ≈98% - — skill only collects and submits, producing default private + noindex + no deployment Each image ** is taken from the photo capture taken during the acceptance of each project: same viewport, same scrolling position, same animation time. Unless otherwise noted, all frames shown are **(meanAbsDiff 0.00). —— 1.

Overview (machine translated)

Chinese | English

Give a URL and have your AI agent copy the website line by line and prove that the copy is correct.

It’s not about “making a page that looks very similar.” It treats the source site’s implementation as a specification: first, it treats the entire site as read-only evidence, then reconstructs the logic line by line from the compressed code, and finally uses a whole set of automatic acceptance gates to prove that “the replication side and the source station are doing the same thing”—including pixel-by-pixel comparison.

Designed according to the Agent Skills open specification, designed for any agent that supports skills. Cross-runtime is not just a slogan but a test: the validation checklist includes both goals where Claude Code is completed and Codex runs the entire process (Hashgraph VC, 166/166 response bytes are consistent)—the same skill directory can run a complete pipeline for any runtime.

Characteristics
Effect comparison: Source site vs. rerun
Start quickly
How it works
Scope of application
How is this different from “Bak Station”?
Verified websites
Warehouse structure
About copyright
Roadmap
Update the records
Contribution
Permission
Friendly links

Features

Grade Assessment First — Before starting work, first check the target category and whether it can be done; If it can’t be done, just state the reason directly, and won’t just run off and produce waste
Forensic Mirroring — Captures the entire site as a read-only snapshot (file-by-file sha256 ledger, reference closure verification, authenticity check), runs even when offline
Traceable line number reverse — Every line in the replica can point to which line in the bundle of the original station is returned; Bugs and strange spellings are copied without revision
Quantitative Acceptance — Automatically compare across five layers: console / network / DOM / geometry / pixel-by-pixel. Differences must either be removed or recorded, and no blurring is allowed
Source Code Delivery — Rewrite the verbatim ported product into a human-readable project (remove modules, name by evidence, add source traceability notes), copy anywhere, disconnect and run
Zero Dependency Toolchain — 75 Node scripts (51 process scripts and acceptance gates + 14 shared libraries + 10 sourced/inference refactorers). Before sourcing, the entire pipeline does not contain any npm packages
Dead Station Can Also Be Saved — Wayback Save Rescue: Anchor + Time Window Choose a Continuous Moment, Original Byte Completed Standard Mirror, Permanent Hole Accurately Registered; Five Dead Station Rescue Tests (Four Revivals, One Complete L3; One Screen Layer Confirmed Completely Lost, Failed Form Registered)
Even RSC Stations Can Be Refactored — Server Components / Next.js App Router are not issued, but its full output (flight stream) is inscribed into each HTML page as a specification: from this, a constructable Next project is reconstructed and closed with a flight semantic gate. In actual tests, a Next 16/Turbopack blog site has identical 18/18 route semantics, a 144 route restation PASS 144/144; blind reverse engineering of public source code volumes: structure ≈95%, behavior ≈98%

This page was translated automatically. Technical details may differ from the original.

Recommended Tools

Try a different keyword or remove a filter.

Install

npx skillfish add boyang-hu/website-rebuild-skill

New to MCP? Start with the mcpskill.org guide →