Citation Cleaner

Clean AI-generated Markdown locally — with evidence for every provider-specific decision.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Citation Cleaner

Inspiration

AI answers are easy to generate and surprisingly hard to publish. Copying from ChatGPT, Claude, Gemini, Perplexity, or AI Overviews often brings citation markers, hidden footnotes, source trails, tracking links, and provider-specific export artifacts. Existing cleanup workflows either upload sensitive drafts or use broad regexes that can silently remove real content.

What it does

Citation Cleaner turns AI-generated Markdown into review-ready copy entirely in the browser. It protects code, removes deterministic citation and export artifacts, shows an inline diff, and reports every applied rule.

Our Build Week extension, Citation Fingerprint, makes Auto mode explainable. It detects concrete provider signatures, reports confidence and matching evidence, and only enables provider-specific rules when the fingerprint supports them. Generic citation syntax stays generic; users can always select a source manually. The public API returns the same structured detection receipt.

How we built it

The project uses Next.js, React, TypeScript, Vitest, and a Manifest V3 Chrome extension. The cleaner is a deterministic rule pipeline: protect Markdown code, resolve the source fingerprint, apply rules allowed by source and intensity, restore code, and return output plus an audit trail.

Citation Fingerprint was designed and implemented with Codex running GPT-5.6 during the submission window. Codex audited the pre-existing architecture, identified that Auto mode previously enabled every provider-specific rule, designed the fingerprint contract, implemented the detector and multilingual report, added tests, and ran the production and extension quality gates.

Challenges we ran into

  • A citation pattern alone rarely proves which model produced the text.
  • Detection had to remain conservative so Auto mode fails closed.
  • The UI supports nine languages, so the new audit trail could not be an English-only bolt-on.
  • The product promises local privacy while also exposing an optional public API; the documentation must make that boundary explicit.

Accomplishments that we're proud of

  • Auto mode now requires evidence before applying provider-only rules.
  • Every source decision exposes confidence and concrete matching signals.
  • Generic input remains Unknown instead of being over-classified.
  • The same detector powers the browser UI and public API.
  • The existing app, API, PWA, and extension remain runnable and tested.

What we learned

Explainability matters even for a small text utility. A safe automation tool should expose not only what it changed, but why it believed a specialized rule was appropriate.

What's next

User-contributed fingerprint fixtures, signed rule-pack releases, more export formats, and a downloadable cleanup receipt for editorial review systems.

Test it

Open the live app and choose Try Perplexity. The cleanup report shows Perplexity with high confidence and concrete detection evidence. Paste generic text with [1] markers to see the source remain Unknown. No account or API key is required.

Primary Codex /feedback Session ID: 019f71c5-bc9c-7160-9c57-57b1eeec0212

Built With

Share this project:

Updates