Blogs
Product UpdatesLong documentsReliabilityAI ModelsNextDocs

Hundred-page documents that actually finish

5 min
Hundred-page documents that actually finish

Hundred-page documents that actually finish

Earlier this year we wrote about generating hundreds of pages of documents and presentations. Being able to ask for a 100-page report is one thing. Getting all 100 pages, every time, with page 87 as good as page 3, turned out to be a harder problem, and one that produced the most frustrating bug reports we had this year: a document that looked finished and stopped on page 14.

This post explains how long generation works, what we changed in v1.11.1, and how to get the best result when you need a very long document.

How a long document gets made

NextDocs does not ask one model to write 100 pages in one breath. Three roles share the work:

  • A router reads your request and decides what kind of document it is, how long, what structure, and which lane to use.
  • A create model writes the document, section by section, streaming it onto the canvas as it goes so you can watch it arrive.
  • An update model handles the edits you ask for afterwards, one section at a time, without regenerating the rest.

Each role runs on the model best suited to it. The router is small and fast. The create model is chosen by your plan and mode: GPT-5.6 Luna on the fast lane, Gemini 3.8 Flash on quality, Claude Sonnet 4.6 on premium. The ladder post explains the choices.

Where it used to go wrong

A model writing a long document has a budget: NextDocs allows about eleven minutes for a creation and six for an update. Some models write at a steady pace from the first page to the last. Others start fast and slow down as the document grows. On a long document, the slow ones run out of budget, the stream ends, and, this was the bug, the half-finished document was saved as if it were complete. No error, no warning, just a report that ended mid-sentence.

We found it by replaying a real 22-page request from production against every model we use. The pattern was unmistakable once we could see it, and it is now the first thing we test for. Five models went through that replay; two of them did not get a lane because of it.

What changed in v1.11.1

Three things.

A stream that ends early is now a failure, not a document. If a model stops before it is done, NextDocs refuses to save the truncated result and tells you what happened, so you can retry rather than discover the problem in a meeting.

An interrupted run keeps what it finished. This is the other side: if your connection drops or you close the tab during a long generation, the pages that completed are kept, and you can continue from there rather than starting over. Silent truncation is a failure; a deliberate stop is not.

Every model on the ladder has passed a long-document gate. No model gets a lane, however new or however good its benchmarks, until it has finished the 22-page replay within budget. Gemini 3.8 Flash passed within a day of its release and was promoted on 3 September. Claude Sonnet 5 stopped at 16 pages and was not.

Getting the best result on a very long document

  • Use Quality or Premium mode. The fast lane is quick and reliable, but for 60 pages with sections that must stay consistent, the quality lane's model holds structure better. Premium writes the most thorough documents of all, and the longest, so give it the time.
  • Give it the structure. "Twelve sections, each with a summary, findings and a recommendations table" produces a far more even document than "a long report". Long documents wander when the shape is left to chance.
  • Generate the outline first, then the document. For anything over 40 pages, ask for the outline as its own short document, fix it, then paste it into the real prompt. It is the cheapest edit you will ever make.
  • Attach the sources. PDFs, Word files and spreadsheets attached to the prompt keep a long document anchored to facts instead of drifting.
  • Edit sections, not the document. After generation, ask for changes to a section by name. The update model touches that section only, which is fast and keeps the rest exactly as it was.

What this means

A long document from NextDocs now ends in one of two ways: complete, or with a clear message that it is not and why. There is no third state. That sounds like a small thing. If you have ever presented from a report that stopped on page 14, you know it is not.

Generate a long document in NextDocs


The NextDocs Team