About

Why this exists

Code survives the trip into a Word document badly. It arrives as grey Calibri with the indentation half-eaten, and by the time the file has been exported to PDF and mailed around, reading it is genuinely unpleasant. Every existing highlighter assumes you have already separated the code from everything else and can tell it which language to use. Documents do not work that way.

So the interesting problem was never the colouring — that part has been solved for decades. It was finding the code: deciding, line by line, whether a stretch of text is a sentence or a statement, with no labels and often no formatting left to go on.

Why it is not built on an AI model

Because most of the deciding evidence is already sitting in the file, plainly readable, and a language model would have to guess at it from flattened text while discarding the very thing that answers the question. Detection here is deterministic: it reads a whole document at once however long it is, it never rewrites a word of your prose, and it gives the same answer today and next year.

A model is still the better tool for some things, and the button that calls one is there for exactly the cases where it is. The full argument, including where it goes the other way, is on why not just use AI.

Who built it

Highlit Docs is built and maintained by Hemant. If you are reading this after cloning or forking the project, replace this paragraph with your own name and background — a real person with relevant experience is worth stating plainly, and an anonymous tool is harder to trust and harder to cite.

How it is funded

It is free and there is no account system. The long-term plan is a paid tier for the features that cost real work — exporting back to .docx and .pdf with the highlighting baked in, batch processing, and additional languages — while the core tool on this page stays free and unrestricted.

There is no tracking of your documents, because your documents never arrive here to be tracked.

Privacy in one paragraph

Your file is uploaded, held in memory while you work on it, and deleted within twenty minutes. It is never written to disk and never shared. Pressing Refine output with AI additionally sends the handful of lines the detector was unsure about — never the file — to Google’s Gemini API for a second opinion, and does nothing until pressed. The full statement is on the privacy page.

Get in touch

Bug reports, false positives, false negatives and language requests are all welcome. The most useful report is the smallest document that reproduces the problem plus a note on what you expected instead.

Email hello@highlitdocs.com.

Roadmap

• True PDF text replacement — the .docx export already replaces code runs outright; the PDF export currently paints over them, leaving the old text in the file underneath. Closing that gap is next.
• More languages — Go, Rust, SQL and shell are the leading candidates. Each is one fingerprint table plus one rule list.
• Inline snippet detection — recognising code mentioned inside a sentence without breaking the paragraph.
• Batch processing — several files at once.