Most bad documents aren't badly written.

Anyone who writes for a living knows this evening. The deck is due tomorrow and forty slides exist that look good: the boxes are aligned, the colours match the brand, someone has spent real hours on it. Then at eleven someone senior reads it end to end for the first time and says: I don't think this is the story.

Nothing is wrong with the slides, and every sentence is competent. The problem sits three levels above them. The argument those slides were built to carry was never agreed, so forty slides of careful work have to go to fix what amounts to a paragraph of thinking.

I have watched this happen to press releases, board memos, white papers, and my own essays. It is almost never a writing problem but an ordering problem. The work got done back to front, so the cheap decisions were made last and the expensive ones first.

Software architects hit the same wall years ago. Their answer was the C4 model, a way of describing a system at four zoom levels: context is how the system sits in the world, containers are the big moving parts, components are what is inside each part, and code is the implementation.

The value is not the diagrams but the discipline the levels impose. You do not argue about a class name while the system boundary is undecided, because the boundary decides whether that class should exist at all. Each level has to be settled before the one below it is worth drawing.

A document has the same shape. So: D4.

Each level has one job, one test that tells you it's finished, and one kind of output. You go down a level only when the level above it survives its test.

Easier shown than described, so here is this article at each of its own levels. Step through them.

This is the whole argument for working top down, and it comes straight from engineering: the cost of a change rises with how late you make it.

Change your D1 while it is still one sentence and you have spent thirty seconds. Change it after D2 and you restructure an outline. Change it after D3 and you bin your research. Change it after D4, which is the eleven o'clock rewrite, and you throw away every hour anyone spent making it beautiful. That is most of the hours.

The method does not change with the form, only the vocabulary does.

You do not need software for this: indented bullets are the whole apparatus. Depth on the page maps to depth in the argument, so a structural problem becomes something you can see rather than something you have to feel.

If a D3 point will not sit under any D2 header, you have found a gap. If one section has eleven children and another has one, you have found an imbalance. Both are invisible in flowing prose and obvious in an outline.

The framework earns most of its keep when a document has more than one parent. Teams that agree the core message and the section layout before anyone drafts prose almost never have the eleven o'clock evening. Teams that skip straight to drafting almost always do.

It also lets people come in at the right level without drowning in context. A subject expert can check your D3 without reading your prose. A designer can start on D4 visuals once D2 is locked. An editor does not have to work your argument out of finished paragraphs, because you can just show them the outline.

And it makes redrafting cheap in the place where redrafting pays. You can sketch four different D2 structures in the time it takes to write one good paragraph, and pick the best before committing to any of them.

This is the part that has changed most since I first wrote about D4, and it is why the framework matters more now than it did then.

Language models are very good at D3 and D4. They will draft your supporting sections, give you five versions of a paragraph, squeeze four pages into one, and polish the rhythm until it sings. They are much weaker at D1. Not because they cannot form a thesis, but because your thesis rests on things they do not have: what you have seen, what you are willing to stake, who you are trying to move.

The practical trick is this: your D1 and D2 are the prompt. Hand a model a topic and you get something generic, because you gave away the only decision it could not make. Hand it a sharp claim and a five-line outline and it will produce something that sounds like you meant it, because you did.

There is a real hazard here too. Models are so good at D4 that polish has stopped being evidence of thinking. A weak argument now arrives beautifully dressed, on time, with confident transitions. So the reviewer's job has quietly moved up: stop being impressed by the prose and go read the outline underneath it.

One question per level, and if the answer is no, you are not allowed downstairs yet.

Open a blank document and write these four lines. Answer them in order, and do not let yourself answer one before the one above it is settled.

That is it: no tooling, no template library, no process to roll out. The discipline is entirely in refusing to work at a level before the one above it holds.

It will not make you a better writer overnight. It will stop you spending your best hours on sentences that were never going to survive, and give you a shared language for the conversation that decides whether a document works.

After enough eleven o'clock evenings, that turns out to be most of the game.