What breaks as the text gets longer
On short text a model is close to flawless, and that is misleading. In a study built for this, models met a length instruction completely below 400 words; above 4,000 words most of them fell under 0.4 on the same scale. What works for two paragraphs does not work for fifteen pages — and nothing signals an error along the way.
Three things break, each in its own way. The first is structure: the model continues by leaning on its own earlier sentences, so towards the middle it starts talking about its subject instead of claiming something about it. The second is repetition: every model tested scored between 60 and 72 on redundancy, meaning that in longer text it keeps returning to the same thought in different words.
The third is the most insidious: missing depth. In the study's own words, models "merely reiterate the key points outlined without delving into the underlying reasons". The text covers the topic in form while paraphrasing its own table of contents in substance. This is hard to notice while reading, because every paragraph makes sense on its own.
Why the outline-first workflow wins
The same study compared the two working methods. When the model was given the task in one pass, its combined score ran between 55 and 63. When it first produced a plan and then wrote section by section, it reached 71 to 86 on the same tasks. Roughly twenty points, with the same model and the same subject.
The reason is obvious once you see it: section by section the model has a short task, and short tasks it does well. And the outline does in advance the work the model is worst at — deciding what belongs where and what stays out. This is the same logic that drives the first-draft approach: execution is the machine's job, structure is yours.
The six-step workflow
This order works for a long article, a study, a proposal and an internal document alike.
- Write down in one sentence what you are claiming. Not the topic — the claim. "Hybrid work is not more expensive, the cost just shows up somewhere else." Without this, the outline becomes a list of topics rather than a structure.
- Ask for three different outlines, not one. For the same material. Usually none of the three is right, but the third will contain a section you would not have thought of.
- You finalise the outline. Section headings, each with one sentence about what that section claims. That sentence is the whole point — without it the model will circle the subject.
- Have it drafted section by section. One section per request. Supply the full outline as context so it knows what comes before and after, but ask only for that section.
- Assemble it and hunt the seams. Transitions, repetitions and drifting register surface here. This step is manual work and cannot be skipped.
- Check at the end. Numbers, names, references. In long text this needs to be a separate step, because the error is usually in the middle, where you are paying least attention.
What a good outline contains
Most outlines fail because they are lists of headings. Five things are worth putting in.
- One claim per section. Not "Costs", but "The cost does not disappear, it moves into equipment purchases". That gives the model something to develop.
- One piece of evidence per section. A number, a case, a quote. Leave it empty and the model will supply something, and you will have to take it out again later.
- The section length in words. "About 250 words." A model holds a length precisely over a short stretch and not over a long one — which is exactly why you give it per section.
- What stays out. A short "we are not writing about these" list. It shapes the text at least as much as the headings do.
- What the reader already knows. One sentence. How much explaining is needed depends on it, and without it the model always starts from zero.
The seams where it gives itself away
Read as a whole, section-by-section text tears in four typical places. All four are quick to fix once you know where to look.
- The recurring transition. "As we saw", "as mentioned earlier" — once per section, so eight times in the finished text. Two is plenty.
- The redefinition. The model re-explains the same term at the start of every section, because it does not know it has already done so.
- The drifting register. The fourth section turns out more formal than the first, especially if you started a new conversation in between. Put the style sheet into every request.
- The contradiction. Two sections give two different numbers for the same thing. This is the most common real error, and only reading the whole thing catches it.
What not to do
Three widespread tricks that measurably produce worse results than the plain workflow.
- Do not ask for the whole text in one pass on the theory that you will fix it afterwards. Fixing it at that point means rewriting, which takes longer than having it drafted section by section.
- Do not chunk and merge blindly. In summarisation this was measured: hierarchical merging and incremental updating gave worse results than a single plain request. Chunking helps on the input side — that is what processing long documents covers — but on the output side the outline is the right tool.
- Do not let word count be the only specification. "Write 2,000 words about it" is the fastest route to redundancy. Length given per section is a constraint; length given in one number is an invitation to pad.
A prompt you can copy
This is the section prompt for step four. Put the outline into every call and narrow the request to a single section.
You are writing one section of a longer article. You get the full outline
so you know what comes before and after — but write ONLY the marked section.
The article's claim: [one sentence]
The reader: [who they are, and what they already know]
Full outline: [section headings + one claim per section]
WRITE THIS ONE: [section heading]
This section's claim: [one sentence]
Evidence to use: [number / case / quote]
Length: about [250] words.
Rules:
- do not summarise the rest of the article, and do not set up the next section
- do not redefine anything that appeared in an earlier section
- do not close the section with a summarising sentence
- anything the supplied evidence does not tell you, do not invent: write [missing]
What to take away
The quality of a long text is settled in the outline, not in the prose. Written in one pass, the model paraphrases its own table of contents; written section by section from an outline broken into claims, the same model produces text twenty points better. Write your claim in one sentence, ask for three outlines, give every section a claim and a piece of evidence, and comb the seams yourself.