Site Skill Builder

Build a Claude Code plugin for a website by looking at the site in your own browser. Observe the pages, propose a roster of skills, get them approved one by one, write them so they survive a redesign, and keep them alive as the site drifts.

The default route is your own signed-in Chrome, via claude-in-chrome — so the agent works inside the session you already have, no credentials are handled, and what it sees is what you see. Headless Playwright/Puppeteer is the fallback for public pages that need an unattended sweep; a plain fetch is used wherever it actually works.

Built for sites you have an account with and the right to use with an agent. It does not bypass authentication, defeat bot challenges, or evade rate limits. It classifies every observation before it is committed, and it has a stop rule for the day a site map turns up something that looks like a vulnerability.

The pipeline

Skill	What it does
site-recon-intake	Target, page list or exploratory brief, account posture, constraints, robots/ToS
pick-browser-route	Own Chrome → headless → fetch; proves the route on one page first
observe-page	Walks the pages; writes an observation note per surface
map-site-api	Records the JSON endpoints worth depending on, with a canary field each
propose-site-skills	Ranked candidate roster — stops for per-line approval
write-site-profile	Every volatile fact in one JSON file
author-durable-skill	Writes an approved skill against the durability doctrine
publication-gate	Publishable / private / never-recorded, before every commit
drift-probe	Re-verifies routes, identity, handles and canaries; reports CHANGED
repair-broken-skill	Fixes what the site moved, and climbs the handle ladder while there

Commands: /site-plugin runs the build, /site-drift runs the maintenance loop.

Agents: skill-durability-auditor (audits a draft for brittle handles and silent failure paths), observation-recorder (writes up captured material, no browser).

The two ideas it exists to enforce

Durability. A site skill is a bet that a page will look roughly the same next month. Make the bet on URL grammar, the site's own JSON, structured data and accessible names — never on generated CSS classes, list positions, or pixel coordinates, which break on a window resize with the page unchanged. Assert page identity before extracting anything, because the expensive bug in this domain is a 200 that serves someone else's content. Fail loudly; never return a partial result as if it were complete. → references/durability-doctrine.md

Publication discipline. Most of what you learn mapping a site is ordinary integration detail. A minority is the provider's defensive posture or a live weakness, and that stays in a private workspace or is never written down at all. Every observation is classified before it is committed. → references/publication-policy.md

What it produces

A public plugin repo — profiles/<site-id>.json, skills, sanitised notes, a drift probe — paired with a private dev workspace holding captures, account-bound material and any findings. → references/repo-layout.md

Neighbours

• browser-data-capture — owns the capture layer: HAR analysis, mitmproxy, live-tab observation, OpenAPI generation, per-domain maps, and disclose-finding for coordinated disclosure. Use it for the capture; this plugin decides what to build from it.
• create-claude-plugins — scaffolds and registers the plugin repo itself. This plugin fills a repo; it does not create one.
• userscript-development — same browser-driven authoring shape, different artefact (Tampermonkey userscripts).

Installation

claude plugins install site-skill-builder@danielrosehill

License

MIT — see LICENSE.