Reverse engineering a website

yat:reverse-engineer teaches you how a live site works by taking it apart in your real Chrome, one rung at a time. You predict, it shows evidence, you explain, and at the end you rebuild a small version yourself.

Setup

1. Install the Claude-in-Chrome extension and connect it to Claude Code.
2. Use any env, or ask for a new one ("new env for web experiments").
3. Say: reverse engineer <url>, I want to know how <specific thing> works.

Pick targets you are allowed to study

Public pages, sites you own, or sites where you have permission. The skill asks, writes the answer into the inventory, and refuses otherwise. It observes only: it will not submit forms, buy things, replay requests, or touch credentials. Secrets are redacted from every note. Full rules: harness/guardrails/web-safety.md.

The ladder

Rung	You predict	You see
R1 Behavior	what the page does	screenshots, optional GIF
R2 Surface	how it's built	DOM structure
R3 Wire	what it asks the server	the real network calls
R4 State	where the data lives	read-only values in the page
R5 Why	why the authors chose this	you argue, then compare to the evidence
R6 Rebuild	what the smallest version needs	you build it
R7 Diff	where yours differs	the same probes, side by side

One rung per turn. You never skip the prediction, and the agent never writes your clone's logic.

Pause and resume

Targets live in envs/<env>/reverse-engineering/inventory.md with their rung, the evidence so far, and a specific next move. Say reverse engineering inventory to list them, or resume reverse engineering <target> to continue with no recap.

Good first targets

A search box with instant results, an infinite-scroll feed on a public site, a docs site's client-side search, your own project. Avoid anything behind a login.