Teaching AI agents to reverse-engineer the web

What it does This is an Agent Skill for runtimes like Claude Code or Codex. Hand it a URL and it grades the target, mirrors the entire site as a read-only forensic snapshot with per-file sha256 checksums, then decompresses and reverse-engineers the minified code line-by-line. A five-layer verification pipeline—console, network, DOM, geometry, and pixel-perfect screenshots—proves the rebuild matches the original. Dead sites can be pulled from the Wayback Machine. The final stage restructures the literal transplant into human-readable source that runs fully offline.

The interesting bit The author treats the original implementation as the specification, not a visual reference. The skill enforces six strict disciplines—copying weird minified patterns verbatim, never “fixing” dead code, and forcing the agent to register every intentional deviation—to prevent silent behavioral drift. It even handles React Server Components by treating the inlined flight stream as the spec, and has reconstructed a 144-route Next.js site with flight-semantic gates.

Key highlights

• Cross-runtime skill tested on both Claude Code and Codex using the same directory; a Codex run on Hashgraph VC achieved 166/166 matching response bytes
• Zero npm dependencies during the rebuild pipeline: 66 Node scripts handle mirroring, verification, and source-ification
• Verified on complex production sites including a 1.25 MB custom WebGL engine with 156 shaders, Apple AirPods Pro (565 webpack modules), and Theatre.js WebGL scenes
• Wayback rescue mode for dead sites; five tested recoveries are documented, with one failure at the visual layer logged as a formal failure case
• Reconstructs RSC and Next.js App Router output from flight streams into buildable projects; tested on a 144-route heavy site

Caveats

• Heavy WebGL or custom-engine builds still take one to three days of agent runtime, though smaller sites now finish in a single unattended session
• The skill defaults output to private, no-index, and non-deployed; all copyright and publication decisions are explicitly left to the user
• One Wayback recovery failed at the visual confirmation layer, which the README records rather than glossing over

Verdict Worth exploring if you need to archive, audit, or learn from complex production sites—particularly those with obfuscated bundles, proprietary WebGL, or dead backends. Not your tool if you just want a quick visual approximation.