Turn any doc Claude writes you into a page you can talk back to. Highlight a sentence, box a diagram, or click a component, ask your question right on the spot, and Claude answers it in the margin — pinned there for next time. No API key, no server, no accounts; it runs on the Claude Code subscription you already have.
You've hit this before: you ask Claude for a spec or an architecture write-up, get back a clean doc, and start reading — then the questions pile up. Why base62 and not a UUID? What does "read-through cache" mean here? Is this number still right? Today those questions go one of two places: into a separate chat where you re-paste the paragraph and re-explain what you meant, or nowhere at all. The doc can't answer, so it quietly becomes one more file you skimmed and never fully trusted.
This keeps the question on the doc. You mark the exact spot that tripped you up, type what you want to know, and the answer lands right next to it — and it's still there tomorrow. A dead export turns into something you interrogate until you actually understand it.
A generated architecture doc: a highlighted phrase, a region box dragged over the diagram, and a pin on the "KV store" node — each holds its own Q&A thread. Green ✓ pins are answered, amber ones are still open; the sidebar shows Claude's answer as rendered markdown.
And it rides the flow you already use — there's no new app to open:
- Ask Claude Code for a doc, spec, or notes. It comes back as an annotatable page.
- Read it in your browser. When a question hits, mark the spot — a phrase, a slice of a diagram, or a whole element — and type it. No context switch, no re-explaining.
- Answer. Claude replies on your subscription: automatically through a local helper you start with one double-click, or by pasting a single self-contained block into any chat.
- Keep. Every question and answer is saved in a plain file beside the doc, so it grows into a reference you trust instead of a snapshot you forget.
Everything stays on your machine: draft questions live in the browser, answers in a <doc>-threads.js file you own. Nothing is uploaded.
/plugin marketplace add vendz/annotatable-html
/plugin install annotatable-html@vendz
That's it. From then on, when you ask Claude Code for documentation/specs/notes for your own reference, it builds them as annotatable pages by default.
- Claude Code (this is a Claude Code skill).
- A Chromium browser (Chrome/Edge/Arc/Brave) for the auto-answer launcher; any browser works for the copy-paste path.
- Node.js — only for the optional auto-answer helper (
annotate-bridge.js). The page itself needs nothing. - The auto-answer helper uses
claude -pon your subscription (no API key). It unsetsANTHROPIC_API_KEYfor the child so it never bills the API.
- Claude writes
<name>.html(with the self-mountingannotate.jsengine) into a shared folder,~/.claude/annotated-docs/, plus a<name>-threads.jsanswers file. - You annotate in the browser; questions are saved locally.
- Answers: either paste the exported batch into any Claude chat, or run the tiny local bridge (double-click launcher) which serves the folder and runs
claude -pto write answers back — the page updates itself. - Reload (or it live-updates): pins turn green, answers appear inline.
The engine self-mounts (injects its own toolbar + sidebar, all CSS classes prefixed az- so it never clashes with your page) and re-locates marks across reloads by text phrase, region fractions, or element selector.
Copy the skill folder into your skills directory:
git clone https://github.com/vendz/annotatable-html
cp -r annotatable-html/plugins/annotatable-html/skills/annotatable-html ~/.claude/skills/MIT © 2026 vendz