If everything fits now, why does putting everything in still make results worse?
Long context changed the question from “what fits” to “what belongs”
When the window was small, curation was forced on you. Now that it isn't, the people getting the best results are the ones still doing it voluntarily.
For a long time the practical limit on working with AI was volume. You couldn't paste the whole contract, the whole codebase, the whole research folder — so you chose. The constraint did your editing for you.
That constraint has substantially lifted, and a predictable thing happened: people stopped choosing.
Fitting is not the same as helping
The assumption behind dumping everything in is that more information can only help, since the model can ignore what's irrelevant. In practice it doesn't cleanly ignore it. Relevant material competes with irrelevant material for attention, and contradictions inside the pile get resolved in ways you didn't intend and can't see.
There's a second effect that matters more. Anything you include reads as endorsed. Paste a superseded version of a document alongside the current one and you haven't given the model two options to weigh — you've asserted both.
What the good version looks like
The people getting consistently strong results with long context tend to do three unremarkable things:
- Say what the material is. A line of framing — "the first file is current, the second is last year's for comparison" — costs nothing and resolves exactly the ambiguity that causes the worst failures.
- Remove the superseded. Not because of space. Because of endorsement.
- Ask for the reasoning path, not just the answer. With a large pile of input, which part of the pile drove the conclusion is the thing you most need to check.
The real shift
Long context didn't remove the editing work. It moved it from a hard limit you couldn't avoid to a discipline you have to choose. Most of the quality difference between two people using the same model on the same material is now sitting in that choice.