Nowely Three decisions, in order, each cheap to change here and expensive to change later. The output is a skeleton: section titles, what each is for, what each deliberately leaves out, a word budget, and the rules that will gate the writing. No prose.
Then it stops. Putting a skeleton in front of the person before two thousand words are written against it is the point, not a courtesy. Measured on 2026-09-11: a ten-section draft written at ordinary quality was abandoned by its reader at the third section, and nine of his nine objections were about what the document contained, where it sat, or how much of it there was. None was about phrasing. Every sentence in it was written against a shape nobody had agreed.
The method, with the measurements behind each stage: stages.md.
Step 1. What comparable documents already solved
A fan-out, not one reader. One agent searching for good examples returns the genre's folklore; six agents on six slices return a sample.
Announce the count and the models before spawning, and wait for the user's word. Default slices, one surveyor each: the exact genre, the same structural position, the most used regardless of genre, vendor guidance, whatever this document's hard part is, and one slice whose job is what not to copy.
Two rules the surveyors carry, both learned by getting them wrong: fetch, do not recall — every document reported carries its URL and its headings in order — and weight by use, not by taste.
Presentation is surveyed here too, from the markdown source rather than from a rendering, because a summary of a document does not show you its devices.
One synthesis decides what to take. The bar is that a change earns its words in this document: it names what it displaces, or admits the document grows.
Step 2. The words
Do not inherit a project's vocabulary because the project uses it. For each load-bearing term: who parses it and as what, where it comes from, and whether its commonest sense in the reader's own field is a different thing.
Where the document's thesis is that A is like B, call A by B's word. Any other choice is an argument against the document, made in every sentence. The worked example — a project that called its delegated agents "seats" while claiming they were the equal of native subagents — is in stages.md.
Step 3. The structure
About ten structures, each from a different reading of what the document is for, not ten runs of one prompt. Announce the count and the models, and wait.
Critics see all of them at once, because ranking is the judgement being asked for and it cannot be made from isolated reviews. Give each critic a different lens and require a fatal flaw for every structure including the one it ranks first.
Ask each critic one more thing: what all of them got wrong. That answer is usually worth more than the ranking — a failure every angle shares is a failure of the brief. On the run this method came from, it was the three critics' shared answer that found the real defect, and none of the ten proposals had.
Synthesise from the winner, grafting only what the critics named. Do not average ten structures into a compromise.
Step 4. What you hand over
One file, and it describes the document rather than arguing for itself. A skeleton introduced by why each section exists — "the block you asked for", "the one you said was missing" — is a negotiation transcript, and a reader feels it before they can name it.
- each section: title, one sentence of purpose, what it deliberately excludes, a word budget
- the mechanical rules the writing must pass, written so that passing is a fact rather than an opinion
- the terminology decisions from step 2, including the ones you rejected and why
- what was deleted outright rather than moved, and the stated cost of deleting it
- any edit this structure requires in a file that is not the document
Then stop and wait. rewrite starts from the skeleton the user agreed to, and routes back here anything
it finds that belongs to a stage above it.
Reference
- The four stages, the measurements, and the content rules: stages.md.
- Filling the blocks, and the loop: loop.md.
- What the field already says about all of this: prior-art.md.