How to Edit a Long Document With AI Without Breaking It
Make this article actionable
Send the article context into Vife Agent and turn it into a plan, checklist, or draft you can keep working on.
A 42-page annual report lands on your desk with a note from your director: "Can you make this read better by Friday?" You paste the whole thing into a chat window, ask for a clearer rewrite, and get back something that looks confident and is quietly unusable. Section 3 has merged with section 4. The table on page 19 is gone. A revenue figure has changed by a digit. Two defined terms now appear in three forms, and the cross-reference on page 31 points at a heading that no longer exists.
That failure is not a model quality problem, and a newer model will not fix it. A whole-document rewrite asks one request to hold three jobs at once: preserve structure, preserve facts, and change prose. When they are mixed, nothing in the instruction tells the model which parts of the page are fixed and which are allowed to move.
The fix is to separate them. Below is the workflow we use for long reports, policy documents, and anything over roughly 3,000 words, followed by the checklist that catches the damage a fast pass leaves behind.
Why the single-paste rewrite breaks
A long document is not one text. It is a structure (headings, numbering, tables, figures), a fact layer (numbers, dates, names, defined terms, quoted statements), and a prose layer (sentences, transitions, tone). A rewrite instruction touches all three at once, and the output is judged on the layer you were paying attention to — usually prose — while the other two drift.
Long outputs also drift within themselves. The first section keeps your heading hierarchy; by the fifth, headings have flattened into bold text. Terminology starts consistent and ends paraphrased. None of this is visible unless you compare against the original structure — and scrolling does not work at 40 pages.
The release pace makes this worse. Microsoft shipped a family of seven models at Ignite on November 17, including coding models it says are "explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work" (Microsoft AI), and LLM Stats lists OpenAI's GPT-6.2 line on November 17 and Google's Gemini 4 Neon on November 18 (LLM Stats). Your document process should survive every one of those releases, which means the constraints have to live in your instructions and your files, not in a model's habits.
Turn the useful parts into next steps
Vife Agent can convert this guide into a prioritized workflow with tasks, risks, and reusable prompts.
Step 1: Map the structure before you touch a sentence
Open a blank file next to the document and build a structure map. This takes ten to fifteen minutes on a 40-page report and saves hours later.
For each section, record:
- the heading text, exactly as written, with its level and number
- the word count
- every table and figure it contains, with captions
- every cross-reference it makes ("see Section 4.2", "as shown in Figure 7")
- every defined term it introduces
Paste the map into the chat as text, not as a description. You are creating the object you will diff against in Step 5. If the document has a table of contents, copy it verbatim as the spine of the map; if not, the heading list is your spine.
Two things become obvious here: which sections are real prose and which are reference material (appendices, tables of figures, boilerplate) that should never be rewritten, and how heavily the document is cross-referenced.
Step 2: Freeze the facts layer
Before any rewriting, extract every fact into a table with four columns: the fact as written, where it appears, its source, and whether it is locked.
Locked means the number, date, name, or quoted sentence may not change. In practice, almost everything in a report is locked. The rewrite is allowed to change how a fact is expressed, never what it is.
A facts table for a 40-page report typically runs 30 to 60 rows:
- every figure with a unit and a period ("$4.2M in Q3 2026")
- every percentage and every comparison ("up 12% year over year")
- dates, fiscal periods, and version numbers
- people, teams, products, and legal entity names
- defined terms and their exact definitions
- direct quotes and attributions
Then add one rule to every subsequent instruction: no number, date, name, or quoted sentence may appear in the output unless it appears in the facts table. Anything else is a flag, not a rewrite. This one rule removes most of the risk people fear when they hand a long document to an AI system: it gives the rewrite a closed vocabulary for the parts that matter.
Step 3: Write one style block and reuse it verbatim
A style block is a short, fixed set of rules that travels with every section request. Keep it in a file and paste the identical text each time. Ten lines is enough:
- Voice: third person, active, no "we" unless the original uses it.
- Tense: present for policy, past for results.
- Sentence length: average under 22 words; split anything over 35.
- Paragraphs: one idea, three to five sentences.
- Terms: use the defined term on first mention, the short form after.
- Numbers: keep the original format and unit exactly.
- Banned: "leverage", "robust", "seamless", "in today's fast-paced", em-dash-heavy asides.
- Lists: only where the original had a list; do not convert prose to bullets.
- Length: within 15% of the original section length.
- Output: the rewritten section only, plus a short list of what you changed.
Rule 9 matters more than it looks. Without a length band, a "clearer" rewrite routinely returns 60% of the original, and the missing 40% is usually the qualification and caveats that made the document defensible.
Step 4: Rewrite one section per request
Now the actual work, and the part people skip. One section per request, with the structure map, the style block, the relevant slice of the facts table, and the original section text. Never the whole document.
The section instruction template
Use the same shape every time:
Rewrite the section below. Keep the heading text and level exactly. Keep every claim; do not add claims. Use only facts from the table. Follow the style block. Stay within 15% of the original word count. Return the section, then list any sentence you could not rewrite without changing a fact.
The last line is the useful one. It turns silent edits into visible ones: the model has to tell you where the constraint bit, and those notes are where you look first during review.
Work in document order. When a section ends with a transition into the next one, rewrite the pair together, because transitions are the first thing a section-by-section pass breaks.
Keep every output in a working file with the section number as a heading. Do not paste over the original: you need it intact for Step 5, and you will want to fall back to it for at least two sections.
Step 5: Reassemble, then diff the structure
Paste the sections into a copy of the document, in order, under their original headings. Then compare against the structure map, mechanically:
- heading count and levels: same number, same order, same text
- tables and figures: same count, same captions, same position
- cross-references: every "see Section X" still points at a section that exists and covers what it claims
- defined terms: each one still defined at first use, and used consistently after
If the original is a document file, a plain text diff against the new version surfaces the rest: dropped sentences, merged paragraphs, vanished list items. Read the diff, not the document. At 40 pages, the diff is the only review that covers the whole text.
Step 6: Run the numbers and names pass
This is a search job, not a reading job. For every row in the facts table, search the new document for the exact string. Then search for near-misses: the same number with a different unit, a slightly different percentage, a name spelled two ways.
Three outcomes are possible, and each has a response:
- Present and identical: pass.
- Present but reformatted (unit moved, rounding changed): fix it back to the original format.
- Absent: either the sentence was cut, or the fact was paraphrased into something unverifiable. Both need a decision, not a silent pass.
Also search for numbers that are not in the table. Any new figure in a rewritten document is either an error or an invention, and both are worse than a clumsy sentence.
Step 7: Do the read-aloud pass
Read the headings in sequence, out loud, as an outline. If the outline no longer tells a story — if section 6 assumes something section 4 stopped saying — the rewrite changed the argument, not just the prose. Fix the argument, not the sentence.
Then read one section aloud in full, chosen at random rather than the one you worry about most. Rhythm problems, repeated openings, and terms that shifted meaning show up immediately when spoken and are nearly invisible when skimmed.
The pre-send checklist
- Structure map exists, and the final document matches it line for line
- Facts table is complete, and every row has been searched in the final text
- No number appears in the final text that is not in the table
- Heading text and levels are unchanged, including numbering
- Every cross-reference resolves to a real section
- Defined terms are introduced once and used consistently
- Each rewritten section is within 15% of its original length
- Tables, figures, and captions are present and correctly numbered
- Banned-word list is applied, and the style block was not quietly abandoned halfway
- A diff against the original has been read, not skimmed
Common mistakes
Rewriting the appendices. Reference material is not prose. Rewriting it creates work and risk with no reader benefit.
Letting the model choose the section boundaries. It will merge short sections and split long ones, and your numbering will not survive.
Treating the style block as a suggestion. If it is not pasted verbatim into every request, the voice drifts by section three, and the drift is invisible until you read the whole thing in one sitting.
Skipping the facts table because the document is "just internal". Internal documents get quoted in external ones. The table is cheaper than the correction.
Reviewing by scrolling. Forty pages of prose cannot be reviewed by eye. Review the diff and the search results.
What this workflow actually buys you
It converts an open-ended request ("make this better") into a bounded one with a defined input, a fixed output shape, and a mechanical check. The model does the sentence-level work it is good at; you keep the structure and the facts, which is where the damage happens.
The artifacts also make the next pass cheap: when the report is revised next quarter, the map and the table are already there, and the diff tells you what changed. If the finished report also has to become a deck or a one-page summary, the map is the outline and the table is the source, so the shorter version inherits the discipline instead of starting over.

