Stop stuffing the context window. In 2026 the real skill is deciding what to leave out.

Models shipped million-token windows this year. Then the research showed they quietly get worse the more you feed them — and that flips how you build.

There's a move almost every maker makes the first time they build something on top of an AI model, and I made it too. The window is huge now, so you fill it. All the docs. The whole conversation history. Every tool you might conceivably need. The reasoning goes: the model is smart, more information can only help it, so give it everything and let it sort things out.

It turns out that's close to backwards.

Chroma published a study in July 2025 that put a name to something a lot of us had felt but couldn't articulate: context rot. They tested 18 frontier models — the GPT-4.1, Claude 4, Gemini 2.5 and Qwen3 families — on how well they held up as the input got longer. The finding is uncomfortable if you've been dumping everything into the prompt: accuracy degrades well before the window is anywhere near full. A model advertised at a million tokens can start dropping answers at a tiny fraction of that. And it's not random. Follow-up work found models lean on the most recent tokens first, then the middle, and treat the earliest stuff in the window worst of all. The window is a number on a spec sheet. Attention is the thing you're actually spending, and it's a lot smaller than the number.

This is why the vocabulary shifted this year. We used to call the job prompt engineering — how do I phrase this. Now the serious version is context engineering, and Anthropic's own guidance frames it well: the quality of an agent depends less on the model than on how you structure everything it sees at inference time. The prompt is one line in a much bigger decision about what makes it into the window at all.

Here's what that actually means for a solo maker, stripped of the enterprise language it usually comes wrapped in.

Treat the window as a budget, not a bucket. Every token you add competes with every other token for a fixed pool of attention. A paragraph of just-in-case background doesn't sit there harmlessly — it dilutes the stuff that matters.

Curate the tools hard. Anthropic has a line I keep coming back to: if a human engineer can't say for certain which tool to use in a situation, the model won't do better. Every tool definition you leave in the context is tax you pay on every turn. Cut to the minimum that does the job.

Retrieve, don't dump. Pull the relevant slice of history or docs for this turn, not the entire transcript. Send everything and let the model figure it out is the expensive habit context rot punishes most.

Put the thing that matters most near the end. Given the recency bias, the instruction you most need obeyed should be the last thing the model reads, not buried in a system prompt from forty turns ago.

I'll be honest about where this hit home for me. A while back we made a decision at Murror that looked, on the surface, like a product or privacy call: we made the AI forget most of what you'd told it every 30 days. We expected it to make the reflections worse, because less history should mean less to work with. It made them better, and people opened up more. I didn't have the words for it then, but that was a context-engineering decision. We weren't throwing away information. We were refusing to drown the model in stale context so the recent, relevant stuff could actually land.

That's the reframe I'd offer. In the era of enormous windows, the instinct to add more is the easy one and usually the wrong one. The harder, more valuable skill is knowing what to withhold. The maker who can look at a full context window and confidently delete half of it is going to ship something that works more reliably than the one who keeps stuffing it — and reliability, not raw capability, is what your users actually feel.

Every token you add is a token competing for attention. Spend them like they're scarce, because to the model, they are.