Why is so good at editing documents?

01

LLM only sees text, not formatting

Open a contract in Word and you see a page: a heading in the house style, a bold amount, an underlined deadline. The Word program creates these formatting, so we can edit visually. Underneath a Word file (.docx) is a complex system of XML, where every word we see is painfully annotated with hidden words known as markup.12

A LLM is not aided by Microsoft Word, LibreOffice or Pages and must rely on its own understanding of the markups. To edit correctly, a LLM needs to understand the effect of the original markup and write the new sentence in the correct markup.

02

LLM output is inherently unstable

Models may have improved substantially but even the latest still make mistakes. In a 2026 Microsoft Research study of 19 models across 52 professional domains, frontier models such as Gemini 3.1 Pro, Claude 4.6 Opus and GPT 5.4 corrupted an average of 25% of document content by the end of long editing workflows.4

In a Word file, one small error is enough to make a file unusable. Here the model has dropped a single slash:

XML has no tolerance for bad markups. A file is either well-formed or it is not. Word can sometimes recover a broken file, but the recovered copy may be missing text or formatting.

Longer output, more mistakes

The risk grows with length. LongProc, a 2025 benchmark of long structured output, found open-weight models faltering at 2,000 output tokens and GPT-4o degrading significantly at 8,000.5 Another study found that performance drops by 13.9% to 85% as the input grows, even when the model can retrieve every relevant fact.6 A 10-page Word document is about 10,000 tokens, trying to edit all 10 pages at once without error is a moonshot.

03

How solves it

We write custom programs that separate formatting from content. The model receives the words it needs to change, together with a rendered image of the page, so it gets the assistance from a Word program. Our code then puts the edited words back into the original runs and leaves the rest of the file untouched.

You will not find this layer in a general-purpose chatbot. The extraction, rendering and reassembly programs are built and maintained by ’s own engineers.

04

Solving the long-context problem

Some edits are long: a paragraph that runs across several pages, or one change applied consistently across hundreds. For these we learned from software engineering. In March 2026, Anthropic accidentally published the source code of Claude Code, its AI coding agent, inside a public npm package.7

That leaked code editor teaches a plain lesson: to edit something long, break it down. We split a document into groups of pages, a technique called batching, and split long paragraphs into sentences, so each edit is small enough to get right.

Breaking a document apart creates a new problem: context. How do you edit a paragraph on page 60 using a definition on page 2? Here, two years of experience in AI code editor (Cursor, GitHub Copilot, OpenAI Codex) pays off. Their answer is context engineering, which gathers related information into a focused collection and give it to sub-agents with clean context windows.8 We do the same with documents. We group logically related information, as a careful reader would, then send it with the edit target to sub-editors that make the actual change.

Example: redacting a judgment

A lawyer wants every witness name in a judgment redacted, and a list of the cases it cites for a junior to research.

1. Search. finds every personal name and every case citation with ordinary pattern matching. No model is needed yet.
2. Classify. Each name, with the sentence around it, goes to a model that decides: witness, lawyer or judge. The answers build a collection of witnesses.
3. Collect. The same method builds a collection of cited cases.
4. Act. Because we know exactly where each witness name was found, we redact those places, and only those, in the original file.

05

Can chatgpt.com, claude.ai and other chat websites do the same?

Sometimes, yes. These websites often include a Python interpreter, which lets the model open a Word or PowerPoint file and edit its XML with code.

But the code is still written by an unstable model, and nothing checks the result against the original. In our internal testing, we often received documents with unexpected deletions and edits. In one extreme case, a 20-page document came back as 10 pages, with no warning.

A silent omission or modification is fatal for any serious work. On , every change our AI agent makes is recorded. You can review each one, revert any of them, and generate a new document.

06

Can computer-use agents like OpenClaw or Manus do this?

These agents control a computer the way a person does. They open the document in Microsoft Word or PowerPoint, then scroll, click and type.11 Because they use the same programs as a human, they can edit like a human, and the results are often impressive.

The price is cost and speed. The agent reads a document by taking screenshots as it scrolls, and every small step needs a model call to decide what to do next. A full-HD screenshot costs between about 1,500 and 2,700 tokens on its own.10 On the OSWorld benchmark, even the best agents took 2.7 to 4.3 times more steps than necessary, and later steps took up to 3 times longer than early ones.9

By our estimate, a computer-use agent can use up to 100 times more LLM tokens than for the same edits, and take far longer to finish.