recreg

Interacting with content is not consent to future recommendations.

---

What it does

You write down what you don't want to see: a topic preset (gambling, alcohol, shopping prompts, distressing news, common phobias) or your own words and phrases. As recommendations load, recreg checks every card against your rules and covers the matches before they render. One click reveals any item, and a "Why?" button tells you which of your words matched.

• Works on YouTube, Reddit and Pinterest (home feed, Up next, Shorts shelf, subreddit feeds, pin grid).
• Only recommendation areas are filtered. Your own searches pass through untouched.
• Optional on YouTube: a Not interested on YouTube button on each covered card, or an auto mode that reports every filtered video to YouTube as you scroll so its own Undo notice replaces the recreg cover. Both are off by default and use your signed-in account.

Install

recreg is not on the Chrome Web Store yet. Loading it takes about a minute on Chrome, Edge, Brave or Arc.

1. Download the extension zip from the landing page, or build it yourself:

   npm install
   npm run build:ext        # output: extension/dist/
   

2. Open chrome://extensions and turn on Developer mode.
3. Click Load unpacked and pick the unzipped folder (or extension/dist/).
4. Open recreg's settings (toolbar icon → Open settings), enable the sites you want (the browser asks permission one site at a time), then add words or apply a preset.
5. Reload any tabs that were already open.

The toolbar badge shows how many items are hidden on the current page. The popup gives you pause, quick-add, per-page status and the YouTube auto toggle.

How it works

your words / presets  ──▶  chrome.storage.local  (never leaves the device)
                                   │
                                   ▼
page loads ──▶ content script finds recommendation cards (site adapter selectors)
                                   │  reads title / channel / aria-label text
                                   ▼
                         matching engine (pure function)
                   whole-word · phrase · substring · Unicode-aware
                                   │
                     match ────────┼──────── no match
                       ▼                        ▼
            cover (Show · Why?)            shown as normal
            or remove entirely

1. Content scripts are registered per site, only after you grant that site.
2. A MutationObserver keeps up with infinite scroll and in-page navigation.
3. Cards stay invisible until they have been checked, so nothing flashes before filtering.
4. Your terms are always treated as literal text, never as regular expressions.
5. If a site redesign breaks an adapter, the popup and badge show a visible degraded state instead of quietly showing an unfiltered feed.

The landing page demo runs the exact same engine the extension ships. Full design and threat model: ARCHITECTURE.md.

Evidence

We measured the presets on 200 video titles we wrote and labelled ourselves (100 for tuning, 100 held out), using the same matching code the extension ships. The numbers below are from the tuning set, after adjusting the preset words on it, so treat them as optimistic.

Preset	Caught	Hidden by mistake
Gambling	11 / 12	3 / 88
Alcohol	12 / 12	1 / 88
Shopping prompts	8 / 12	4 / 88
Distressing news	11 / 12	1 / 88
Common phobias	9 / 12	4 / 88
All	51 / 60 (85%)	13 / 440 (3%)

What it gets wrong, on purpose shown: Spider-Man: No Way Home is covered (matches "spider"), Casino Royale film review is covered, HUGE sp1der found in garage slips through, and titles with no keyword at all are missed.

Limitations of this measurement: titles only (no thumbnails or descriptions), all in English and written by us, small samples (one title moves a preset by about 8 points), human labels. The evaluation kit and raw results are in recreg-eval/; node recreg-eval/measure.mjs --split dev reproduces the table.

Privacy

Private by architecture, not by promise.

• No servers, no accounts, no telemetry. Your filter list lives in chrome.storage.local on your device. There is nothing to leak to.
• Zero network access. The extension requests no network capability at all; the manifest asks only for storage and scripting. A CI gate fails the build if fetch, XMLHttpRequest, innerHTML or eval appear in src/.
• Least privilege. Site access is granted per site, by you, only for the sites you enable. No <all_urls>, no tabs, no webRequest.
• Your filters are not displayed on the page. Covers never name the matched word unless you press "Why?".
• YouTube feedback is opt-in. The only thing that ever reaches a platform is YouTube's own "Not interested", sent through its own menu, and only if you turned that on. recreg never fills in or submits "Tell us why".

Development

npm install
npm run verify          # security gate → typecheck → tests → build
npm run dev             # build the site and serve it at http://localhost:4173

Script	What it does
npm run verify	security check → typecheck → tests → full build
npm run build:ext	extension → extension/dist/
npm run build:frontend	landing site + extension zip → frontend/dist/ (what Vercel deploys)
npm test	engine, preference-validation and adapter unit tests
npm run check:security	fail if network calls, HTML sinks or dynamic code appear in src/

src/engine/       pure matching engine, presets, preference schema (no DOM, no I/O)
src/adapters/     per-site selector definitions — data, not code
src/extension/    MV3 extension: service worker, content script, popup, options
src/site/         landing-page scripts (reuse the engine)
frontend/         landing page HTML/CSS/assets (Vercel root is frontend/dist)
recreg-eval/      preset evaluation kit and labelled titles
tests/            unit tests
scripts/          security gate, static dev server
build.mjs         esbuild bundling, icon generation, extension zip

Known limitations

• Selector drift: YouTube, Reddit and Pinterest change their DOM regularly. Adapters in src/adapters/index.ts carry fallbacks but need periodic checking against the live sites.
• Word matching is literal: a rule for bet does not match bets. Presets include common inflections; add your own as needed.
• Text only: image-only content with no text or alt labels can't be matched.
• Instagram is not supported: its obfuscated, fast-changing DOM makes a reliable adapter a project of its own.

Roadmap

• Signed remote adapter updates, so site redesigns don't wait on store review. This would be the only network feature the extension ever gains.
• Firefox build (same WebExtensions codebase, minor manifest differences).
• On-device semantic matching on top of keyword rules.
• More sites: X/Twitter, TikTok web, news aggregators.