rewrite

AI models speak Markdown. People read documents.

rewrite bridges the gap. Markdown carries the content from one model to the next; you get to read and work with it as a document.

Open it in your browser and start working — no account or API key required. Installing it as an app is entirely optional.

Try it. The sentence below is deliberately clumsy: it repeats two expressions and makes the same point twice. Select it and ask rewrite to improve it.

This is a very simple sentence that you can select to see concretely how rewrite can concretely improve a very simple sentence.

Work on one passage at a time

Select → choose an action → use the AI → bring the response back → decide.

rewrite prepares the request for the passage you are working on. Send it to whichever AI you prefer using the clipboard, or run it directly on your device when Chrome makes its built-in model available.

You can always select the whole document. Most of the time, a focused request produces a better answer and lets you review the result without rereading pages of text to work out what changed. You decide exactly which part of the document the AI may touch.

The seven available actions

rewrite works on the selected Markdown lines. In Document view it always highlights the full range of lines that will be included in the request; anything outside those lines is neither sent nor modified.

Expand, Summarize and Translate each return a complete replacement for the selection. You may apply it manually or automatically; either way, the original remains recoverable from its annotation.

Image is intentionally different from the other actions. It does not revise the text or return annotations. It uses the meaning of the passage as the starting point for a new visual asset, while leaving the complete prompt in your hands.

For anonymization, rewrite detects the most mechanical clues locally — email addresses, telephone numbers, codes and dates. They are candidates, not decisions. The AI reads the context and may add, discard or reclassify them. The selected text is therefore sent to the model you choose: the action prepares a copy for later sharing, but it cannot hide the source from the AI used to anonymize it.

Text rarely starts on a blank page

Writing rarely begins on a blank page. More often, the raw material is an AI response, a set of notes, a web page, a PDF, a Word document, an email, a presentation or research that has not yet been carried out. rewrite treats all of them as starting points for the same working copy: clean Markdown that reads like a document and is ready to be transformed, checked or revised.

Start with material you already have

Copy and paste is the quickest route. You can also import Word documents, PDFs, spreadsheets, presentations, OpenDocument files, web pages, email and EPUBs. Files chosen from your device are read and converted to Markdown in the browser; they are never uploaded to rewrite.

Link import is intentionally limited to Wikipedia articles, arXiv papers and direct public PDF URLs. Wikipedia content comes straight from the public MediaWiki API. An arXiv paper comes from its official HTML representation, with the official PDF as a fallback. A direct PDF is downloaded by your browser when its host permits cross-origin access. Nothing passes through a rewrite server. For any other web page, open it normally, copy the content you need and paste it into rewrite. If a PDF host refuses direct browser access, download the file yourself and import the saved PDF.

The aim is not to reproduce the original layout. It is to preserve meaning, restore the reading order and keep tables intelligible. With a two-column PDF, rewrite reconstructs the sequence of the text rather than imitating the position of every element on the page. Any metadata found in the source stays attached to the working document but outside the text sent to the AI.

PDFs with automatic tables of contents, deep heading hierarchies, tables or elaborate layouts may lose some structure when extracted in the browser. When rewrite detects one of these cases, it completes the import and warns you. If structure matters, the most reliable route is to open the PDF in Word, save it as DOCX and import the DOCX file.

Research from sources: when you still need to find and select the material

Research from sources covers the stage before a document exists: finding verifiable material, choosing a story and turning it into usable Markdown. It is not an eighth editor action, but another way for a working text to begin.

Choose an editorial preset — AI, Cybersecurity, Innovation, Science, Current affairs, Business and markets, Culture and entertainment, Sport, Environment and climate or Space exploration — or create one of your own. Then choose the output language. rewrite builds a research request for an external AI with web access: the model you choose performs the research, not rewrite. Sources may be in any language and may include local reporting, specialist publications and original material from China, Japan, Korea or elsewhere; titles, summaries and the finished article come back in the language you selected. Local AI is deliberately unavailable here because research depends on current web sources. If nothing genuinely meets the criteria, the response may say so rather than filling the requested number of candidates with weak material.

The structured response becomes a table of candidates. Each candidate carries identified sources and compact evidence passages. Choose one story and rewrite copies a complete article request that can be pasted into a new conversation or given to a different model: it does not rely on hidden chat memory. It asks for clean Markdown, a proper H1 title and descriptive source links. When the article comes back, rewrite removes an unnecessary outer Markdown fence, turns standalone URLs into readable links and locally compares every link with the sources declared for the selected candidate. Unmatched links are not removed: they remain flagged in the document metadata. Here, source links are a deliberate part of the article rather than import metadata.

Research and drafting no longer have to take place in the same conversation. The second request explicitly includes the selected candidate, its sources and its evidence passages, so you may keep the same model or deliberately switch providers. Portability is not independent verification: the finished article still opens directly in a new tab, where you can immediately run Fact-check with another model. Research from sources extends where a document can begin, not how it is reviewed: from that point on, the article is an ordinary rewrite working copy.

The workflow ends without turning into an archive. rewrite retains only the story you actually used, together with its sources, in a local memory attached to the preset so that the same story is not suggested again. The resulting tab also records the preset, the number of alternatives returned, the selected story and, when the response declares it, the model that performed the research. Rejected candidates and the conversation itself are not retained.

What this passage actually says

The seven actions transform, check or create material. This does none of those things. It reads.

When the local model is available and understands the selected language, a ? appears next to Actions. It does not rewrite the passage or alter the document. Instead, it explains in plain language what the text claims, what it takes for granted, where it is ambiguous and which questions it leaves unanswered.

This is a check that usually happens too late. If the model's reading does not match what you meant, the problem is already in the paragraph — and you find out before your reader does. The task is deliberately narrow and testable: understand the text, not rewrite it to suit a preference.

This reading tool is available only through local AI. If the browser model is unavailable or does not support the selected language, the button does not appear.

Bring the answer back into the document

This is the step most tools leave out. Copy the AI's answer — using Copy Markdown when your AI provides it — return to rewrite and choose Read clipboard. With local AI, the result returns directly through the same workflow.

rewrite reconnects each suggestion to the passage it came from, even if you edited the document while the AI was working. Suggestions appear as margin annotations. Review them one by one, or enable automatic application separately for each action. Recognized changes then enter the text immediately, while the original remains recoverable from its annotation.

The option is not shown for Fact-check, which does not modify the text, or Image, which creates a new tab.

If a suggestion cannot be anchored — because the AI misquoted the source or the passage no longer exists — it never disappears silently. It stays with the document until you decide what to do with it.

Two views of the same text

When ##, ** and table pipes are visible, you are reading syntax as well as prose. That makes the rhythm of a document harder to judge. Document view shows the text as a reader will encounter it; MD view exposes the source. Both are views of the same text, not separate copies.

Change model without changing tool

Serious work rarely relies on a single AI. One model may write better than another, and a fact-check is more useful when it comes from a model that did not write the passage in the first place.

Copy and paste may look like a workaround, but it is the one protocol every model already supports. Use the subscription you have, choose the model you trust and avoid integrations that need approval. rewrite prepares the passage, keeps track of what comes back and never locks the document to one provider. Local AI simply removes the clipboard step when the browser can do the work itself.

What stays in the browser

A website cannot freely access files on your disk. Here, that limitation is useful: files enter only when you choose them, and results leave only when you decide. By default, documents, annotations and the workspace remain in the browser. When you sign in to the private service, rewrite also keeps an authenticated-encrypted server copy of the workspace until you delete it. URL imports are retrieved directly by your browser from the public source; rewrite does not act as an intermediary.

The local model is supplied by Chrome, not rewrite, and inference runs on your device. rewrite cannot make promises about Chrome's surrounding behavior beyond the information published by the browser vendor.

The workspace saves itself and can be exported as a single file. rewrite is not trying to become your document archive. It is a temporary, local workbench. The finished document still belongs where you normally keep it — Word, Drive, SharePoint or a document management system.

Use the link or install the app

The link remains the simplest way to use rewrite. If you prefer a dedicated window and a system icon, open Application in Settings. Chrome and Edge can install it directly; Safari uses Add to Dock. The installed app continues to run on the browser engine that created it.

Once rewrite has loaded, its interface and essential local converters remain available offline. Importing from a URL — or doing anything else that consults an external resource — naturally still requires a connection.

Local AI in the browser: a glimpse of what comes next

Today, the clipboard is the universal bridge to any AI. The next step is different: the model lives in the browser, receives the selection directly and returns its answer to the same document — with no extra account, API key or intermediary service.

rewrite already treats this as a native editorial workflow, not a separate technology demo. The selection, presets, suggestions, annotations, application and undo behavior stay the same. Only the route to the model changes.

When Chrome exposes its built-in AI and the local model has been downloaded, Try local AI appears alongside the clipboard route. It is available for Improve, Expand, Summarize, Translate, Anonymize and image-description preparation. Fact-check is deliberately excluded: meaningful verification requires a model that can consult current sources.

Chrome's Prompt API — the browser interface that lets websites use its local model — is available only on compatible desktop installations, not on smartphones. Before offering the local route, rewrite checks every language the operation needs: the selection, the instructions and the response. If the installed model cannot handle any one of them, the clipboard route remains available.

According to Chrome's documentation, compatible releases from Chrome 149 support English, Spanish, Japanese, German and French for both input and output. Italian is not yet on that list. rewrite still checks the browser for every operation rather than assuming support, because languages and capabilities can change from one release to the next.

The reading tool described above also relies entirely on this local route. That is why it appears only when the model is genuinely available and understands the selected language.

While the model is working, the selected passage stays inside the browser workflow. There is no response box to fill and no need to leave rewrite. The answer returns through the same suggestion, annotation and replacement system used for clipboard responses.

Long translations are handled in chunks. rewrite divides the Markdown at natural boundaries, translates each block in sequence and shows its progress. Each block gets a fresh session; if it still exceeds the model's limit, it is split and retried. The original document is changed only after every block has completed successfully.

When both routes are available, rewrite asks which one you want. Choosing local AI leaves the clipboard untouched. Choosing another AI copies the prompt only then. Local processing can be stopped, and rewrite checks the request size against the capacity reported by the browser before it starts.

This capability is still evolving. Browser APIs, local models, supported languages and context limits continue to change quickly. A local result may be less accurate than one from a larger model, and some requests may fail. The universal clipboard route therefore remains available.

The first availability check is voluntary and does not trigger a download. Once a check succeeds, rewrite checks again automatically on later visits. It never assumes the model is still present and never downloads it without confirmation.

Between the chat and the finished document

rewrite lives between an AI chat and the document you intend to keep. It handles the material created before the final version and while content moves between models: text extracted from a PDF, cleaned drafts, a paragraph rewritten three times, a fact-check handed to a different AI. The kind of material that would otherwise vanish into an old conversation or a folder full of files called document-2-final-ok.

Input and output formats

Every supported file becomes Markdown, the shared working copy behind every view, annotation and AI-assisted operation. Conversion prioritizes text, reading order and useful structure rather than attempting to reproduce the original layout.

From a file to Markdown

The legacy .doc, .xls and .ppt formats are recognized but not converted blindly. rewrite asks you to resave them as .docx, .xlsx or .pptx.

From a link to Markdown

From Markdown to a file

Saving the complete workspace is a different operation. It creates a restore file containing every tab, its history and associated information, plus the configuration of all seven actions — including presets, glossaries, phrases to avoid and protected elements. Use it to continue working in rewrite, not to distribute the finished document.

A revision that leaves a trace

Writing “reviewed with AI” at the bottom of a document proves very little. What matters is the process: which passages were selected, which automated checks were run, which suggestions came back from the AI, and which of them a person accepted, rejected or left unresolved.

The distinction also has practical legal significance. From 2 August 2026, in the specific case of AI-generated or AI-manipulated text published to inform the public about matters of public interest, Article 50(4) of the EU AI Act provides an exception to the disclosure obligation where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility. The final European Commission guidelines, published on 20 July 2026, explain how these obligations should be applied and confirm that they take effect on 2 August 2026.

A grammar correction documents a grammar correction — nothing more. It does not show which claims, choices or passages were actually examined, and a label at the bottom of the page adds no evidence. What matters is being able to show what was reviewed and which decisions were made.

rewrite provides the tools for that record. It preserves the selection, the requested action, measurements calculated locally, the response received, the suggestions applied and anything still unresolved or no longer anchored to the source. Exporting a commented Word document carries that trail outside the application, where it can accompany the finished work.

The record kept by rewrite can help document the process, but it does not establish regulatory compliance by itself. Editorial responsibility remains yours. When you need to show that a text was genuinely reviewed, you can present the trail instead of merely making the claim.

This text is already inside rewrite

What you have just read is a document open inside the application, not a conventional page describing it from the outside. Switch to MD view and edit it. Your copy stays in the browser.

The r·w logo reopens the published original at any time. If doing so would discard local changes, rewrite warns you first.