WriteWell

Proofread and rewrite text anywhere in Chrome — entirely on your own machine.
Powered by Gemini Nano. No API key, no server, no account, and no network request carrying your text.

Select text on any page, right-click → WriteWell, pick an action. The result streams into a floating card with a Copy button.

WriteWell never edits the page

This is the central design decision, and everything else follows from it. The extension reads your selection and shows you a result — it never writes back into the page. You copy and paste yourself.

That sounds like a limitation. It buys a lot:

• Nothing can be destroyed. No fighting React controlled inputs that revert changes, no corrupting rich-text structure in Notion or Google Docs, no breaking the page's undo stack, no firing input events that trip the page's own listeners. That is where extensions in this category accumulate most of their bugs, and WriteWell doesn't have the category.
• It works on text you can't edit. Polish a sentence in a document you don't own, or on a page you're only reading.
• The permission set stays tiny. See Privacy.

Actions

Action	What it does	API
Check grammar	Fixes spelling, grammar and punctuation, preserving your wording. Shows the original with each correction highlighted, and the corrected text below.	Proofreader
Polish	Improves clarity and flow while keeping your meaning and voice.	Prompt
Friendly	Warmer, more casual tone.	Rewriter
Formalize	More formal register.	Rewriter
Shorten	More concise, without losing essential meaning.	Rewriter

Check grammar is the only action that renders a diff, because the Proofreader is the only API that reports which ranges it changed. The rest return prose:

Requirements

Gemini Nano's hardware bar is high, and a meaningful share of machines will not clear it:

• Chrome 138+ on Windows 10/11, macOS 13+, Linux, or a Chromebook Plus
• ~22 GB free on the drive holding your Chrome profile
• More than 4 GB of GPU VRAM. Integrated GPUs can qualify — Chrome appears to count shared system memory — so don't rule yours out based on the dedicated-VRAM figure alone.
• 16 GB RAM and 4+ CPU cores
• An unmetered connection for the one-time model download, which is several GB

Check before installing. Open any page, press F12, and run:

await LanguageModel.availability()

available or downloadable means you are good to go. unavailable means this device cannot run the model — WriteWell will say so plainly rather than spinning forever. English only for now.

Install from source

Not on the Chrome Web Store yet. To run it:

git clone https://github.com/zeroth/WriteWell.git
cd WriteWell
npm install

Then:

1. Open chrome://extensions and turn on Developer mode (top right)
2. Click Load unpacked and select this directory
3. Open WriteWell's Details → Extension options and click the download button to fetch the model. This has to happen on that page: create() requires user activation while a model is missing, and a service worker never has any.
4. Select text on any page → right-click → WriteWell

npm install is only needed for the tests. The extension itself has no build step and no runtime dependencies — what is on disk is what Chrome runs.

About the trial-gated APIs

The Proofreader and Rewriter APIs shipped as origin trials, and those trials have lapsed. On a stock Chrome they report as unavailable, and WriteWell falls back to the Prompt API automatically — all five actions keep working with somewhat weaker output, and Check grammar returns a plain rewrite instead of a highlighted diff.

To get the real Proofreader locally, enable chrome://flags/#proofreader-api and relaunch Chrome. That is currently the only way to exercise the diff.

How it works

right-click / keyboard shortcut
        │
        ▼
service worker ──── injects the content script via activeTab
        │
        ▼
content script ──── reads window.getSelection() itself,
        │            never info.selectionText (Chrome truncates it)
        ▼
   long-lived Port
        │
        ▼
service worker ──── owns every AI session
        │
        ▼
    ai-client  ──── picks the API, streams deltas back
        │
        ▼
      card     ──── shadow root: renders, copies

Two constraints shaped this:

Inference runs in the service worker, reached over a chrome.runtime Port rather than one-shot messages. Port traffic resets the MV3 idle timer, which matters because a long generation would otherwise get the worker killed mid-stream.

Onboarding exists because of user activation. create() demands it while the model is undownloaded, a service worker never has it, and it cannot be forwarded between contexts — so the download must be started by a real click on an extension page. After that first run the model is available browser-wide and the worker creates sessions freely.

Project layout

manifest.json                MV3; five permissions, no host_permissions
src/background/
  actions.js                 the five actions, as a data table
  ai-client.js               the ONLY file that names Chrome's AI globals
  service-worker.js          context menus, keyboard command, port routing
src/content/
  content.js                 the only file that touches the host page
  card.js / card.css         shadow-DOM result card
  selection.js               selection reading, viewport clamping
src/onboarding/              setup page; doubles as the options page
src/shared/
  protocol.js                message types shared by both sides
  corrections.js             Proofreader ranges → highlight segments

Adding a sixth action is a row in actions.js. Every trial-gated action carries a Prompt API fallback there, which is why a lapsed origin trial degrades output instead of breaking the extension.

Development

npm test

npm run test:watch

Requires Node 22.22+ or 24.15+ — see engines in package.json for the authoritative range. Vitest and jsdom are the only dev dependencies.

After changing anything, press reload on WriteWell's card in chrome://extensions, then refresh any page you want to use it on. A page still running the old content script shows a [WriteWell] console warning telling you to reload, rather than silently doing nothing.

Testing philosophy

Gemini Nano cannot run in CI. No headless runner has a 22 GB profile volume and 4 GB of VRAM, so every automated test here runs against fakes. The suite covers what fakes can honestly cover: correction-range math, action routing and fallback selection, session caching and cleanup, card rendering, viewport clamping, and the port protocol.

Everything a fake cannot reach lives in docs/manual-checklist.md, a written checklist run by hand against a real browser. It is not a nicety — it is the only verification that ever touches a real model, and it has already caught bugs the suite could not, including a missing output-language declaration and a highlight painted over bare whitespace.

If you change behaviour only a browser can observe, add a checklist item. Never tick a box you have not personally performed.

Privacy

WriteWell requests exactly five permissions: contextMenus, activeTab, scripting, storage, notifications.

There is no host_permissions and no <all_urls> content script. No code runs on any page until you explicitly invoke WriteWell there, and activeTab grants access only to the tab you invoked it in. Your text is processed by a model on your own machine and is never transmitted anywhere. There is no analytics, no telemetry, and no account.

Documentation

• docs/manual-checklist.md — the browser checks CI cannot do
• docs/superpowers/specs/2026-08-17-writewell-design.md — the design, and why it is shaped this way
• docs/superpowers/plans/2026-08-17-writewell.md — the implementation plan, including as-built notes where review found the plan's own reference code defective

Further reading

• Chrome built-in AI
• Built-in AI API status
• Proofreader API
• Extensions and AI

License

MIT — do what you like with it, no warranty.