Second Draft

A Chrome extension that rewrites AI-sounding text so it reads like a person wrote it. Select text on any page, review the edits, apply them in place.

It removes em dashes, AI vocabulary like "delve" and "tapestry", chatbot filler, forced rule-of-three lists, and other patterns from Wikipedia's "Signs of AI writing" research.

Rewrites run on your device using Chrome's built-in Gemini Nano. Nothing is sent anywhere unless you add your own API key.

Demo · Privacy policy · Changelog

Version 1.3.0. Not yet on the Chrome Web Store.

Features

• Select text, right click, choose Humanize; the popup opens and starts on its own
• Every edit is listed as old text to new text, with the reason it changed
• A count of tells found and tells remaining, and a check that no numbers, names, dates, or quotations went missing. A rewrite that drops one is redone automatically, once. On the API-key path that second pass costs a second request.
• AI-flavored words are clickable; pick a plain replacement before applying
• Apply the rewrite back into the page, undo it, or try again for a different one
• Keyboard shortcut (Ctrl+Shift+H) and a right-click menu entry
• Add your own phrases to flag
• Upload a writing sample (.txt, .md, .docx) to see your sentence statistics and get warned when a rewrite drifts from them
• Disable per site; paste box in the popup for sites that block editing

Install

Not on the Chrome Web Store yet.

git clone https://github.com/Joel-Wwalker/second-draft.git
cd second-draft
npm install
npm run build

Open chrome://extensions, turn on Developer mode, click Load unpacked, and select .output/chrome-mv3.

Requires Chrome 138 or newer. Download the on-device model from the extension's options page. Without it, the extension uses its deterministic cleanup rules and labels results accordingly.

How it works

1. Rules. Twelve patterns detect AI tells. Four are fixed directly in code (em dashes, curly quotes, emoji, chatbot filler). The rest are reported to the model in the prompt.
2. Rhythm and vocabulary weight. Both measured, not requested, and both calibrated against 1000 human-written paragraphs so the thresholds sit at the edge of human range rather than in the middle of it. Sentence length spread: human median 0.41, machine 0.16, flagged under 0.22. Share of eight-letter words: human median 0.19, machine 0.34, flagged over 0.30. The prompt gets the measured numbers and the offending words, and a rewrite that stays flat or heavy is redone once. npm run compare a.txt b.txt prints the same numbers for any text. Sentence opening variety was tried and dropped: it flagged 57.6% of the human prose.
3. Engine. Gemini Nano on device, or an Anthropic or OpenAI-compatible endpoint if you add a key. Both stream.
4. Enforcement. The rules run again on the model's output. A prompt cannot guarantee there are no em dashes; code can.

The text-processing code is a pure module with no DOM or extension API access. That is why adding three engines required no pipeline changes. Details in docs/ARCHITECTURE.md.

Development

npm run dev         # WXT with hot reload
npm test            # unit tests (vitest)
npm run typecheck   # tsc --noEmit
npm run build       # -> .output/chrome-mv3
npm run e2e         # Playwright against the built extension (build first)
npm run zip         # store upload artifact
npm run icons       # redraw the extension icons
npm run screenshots # re-render docs/screenshots (build first)
npm run compare     # measure texts against the signals the engine acts on
npm run eval        # aggregate a batch of rewrites (see eval/README.md)

TypeScript strict with noUncheckedIndexedAccess, no any, zero runtime dependencies. The .docx reader is a ZIP parser written against the platform's DecompressionStream instead of a library.

Testing

239 unit tests and 2 Playwright tests that drive the built extension in a real browser. CI runs both on every push.

Guards are verified by mutation: break the guard, confirm the named test fails, restore it. Fixtures assert hand-counted values rather than whatever the code returned. This caught a regex that deleted trademark symbols, a chunker that could overflow the model's context, a .docx upload that expanded 199 KB into 200 MB, and a right-click path that read selections on sites the extension was switched off for.

Some of it cannot be tested automatically at all. The page script is injected only when a real right click grants activeTab, and no test runner can produce a native context menu, so replacing text on a live page is covered by the manual matrix rather than by Playwright.

docs/manual-test-matrix.md has 49 rows covering what automated tests cannot: real sites, real models, and browser APIs that do not exist in a test runner.

Project layout

src/
├── engine/        rules, prompts, providers, pipeline
├── shared/        types, diff, storage, profile, docx, sse, redaction
├── content/       selection capture, in-place replacement, undo
├── background/    handing a page selection to the popup
├── entrypoints/   background service worker, popup, options
└── types/         ambient types for Chrome's Prompt API
docs/              privacy policy, store listing, test matrix, architecture
docs/design/       the spec and plans this was built from
docs/screenshots/  generated by npm run screenshots
samples/           texts used to set the style thresholds, ours and a rival's
eval/              how a prompt change gets judged on sixty samples, not one

Contributing

See CONTRIBUTING.md, particularly the testing requirements, and CODE_OF_CONDUCT.md. Report security issues through SECURITY.md rather than public issues.

License

Copyright (c) 2026 Joel Walker. AGPL v3 or later (see LICENSE and NOTICE). Commercial licenses without the AGPL requirements are available: COMMERCIAL.md.

Attribution

Rewrite patterns come from blader/humanizer SKILL.md v2.8.2 (MIT, vendored at docs/skill-source/), based on Wikipedia's "Signs of AI writing" by WikiProject AI Cleanup.