why pasting in everything makes ai worse.

elisabeth hitz · june 18, 2026 · 5 min read

the natural instinct is to give it everything. paste the whole document, include every message, add all the context you can find. more information, better answer, right? not always. and the failure is sneaky, because the output still looks complete.

anyone who has crammed for an exam knows the real limit: there is only so much you can hold in mind at once, and the things in the middle disappear first.

the middle vanishes

psychologists named this a century ago: the serial position effect. items at the start of a list get rehearsed more and stick. items at the end are still fresh and stick. the middle gets neither advantage, so it drops out. run a quick memory test on yourself with a list of fifteen words and you will recall the first few and the last few, and lose the ones in between.

here is the part that matters for your work: large language models do the same thing. in 2023, stanford researchers placed a key fact at different positions inside a long context and measured whether the model could use it. accuracy was highest when the fact sat at the very beginning or the very end, and dropped by more than thirty percent when it was buried in the middle. this is not a quirk you can prompt away. it is structural, attention weights the edges of the window more heavily.

what this means in practice

if you paste a twenty-page document and ask about something on page eleven, the model is more likely to miss it than something on page one or page twenty. the dangerous pattern is the important instruction sitting in the middle of a long thread, where it quietly loses weight. the safer pattern is the same instruction stated up front and restated near the end.

so the practical rule is short: put your most important instructions at the beginning and the end. if a constraint absolutely must be followed, state it early and repeat it late. do not trust the model to give equal weight to everything in between.

the real skill is curation, not volume

this is the core tension of working with context. every piece you add pushes the other pieces further toward the dead zone in the middle. so the question is never just what to include, it is where to put it and what to leave out. curate ruthlessly, place what matters at the edges, and repeat the load-bearing instruction.

which is exactly why two things from the rest of this series work. a tight brief that names the deliverable, inputs, and nuance beats a giant dump, because it is short enough that nothing important hides in the middle. and a skill file beats re-pasting your rules every time, because it puts the constraints in a fixed, prominent place instead of burying them in a long chat. more is not the fix. smarter placement is.