Context Is the Whole Game

The difference between a thin answer and a sharp one is almost never the prompt wording. It is what the model can see. Anthropic calls the discipline of managing that "context engineering."

Last updated · June 30, 2026

The model has no memory of you. Every answer is built only from what's on the desk in front of it, right now.

This took me a while to internalize. The context window isn't a memory — it's a workspace, a desk. Whatever you've placed on it — your instruction, the files, the conversation so far — is everything the model has to work with. Nothing outside that desk exists to it. It doesn't “know” your project; it knows what's currently in front of it. Once that clicked, a lot of frustration made sense: when the model “forgot” something, it hadn't forgotten — it had never been told, or the thing had slid off the edge of the desk as a long conversation grew.

The model reasons perfectly about what's on the desk — and is blind to everything off it. Most “dumb” answers are really an empty spot on the desk.

A worked example

Say you ask Claude to fix a bug and it changes the wrong function. Nine times out of ten the desk was wrong: you pasted one file, but the bug lived in how two files interact, and the second was never on the desk. The model reasoned perfectly about what it could see — it just couldn't see the thing that mattered. Add the second file, and the “dumb” model is suddenly sharp. Curating the desk is the work; the thinking is the easy part.

This is the skill Anthropic calls context engineering: finding the smallest set of high-signal things that make a good answer likely — and no more. Because the desk has a second, less obvious rule: more is not better. A model has a limited attention budget, and as the desk fills with junk — a giant pasted log, twenty messages of dead ends — its accuracy on what actually matters quietly degrades. The clutter doesn't just waste room; it pulls focus off the signal.

Cluttered desk

the whole repo pasted in, plus twenty messages of earlier dead ends, then — somewhere — the question

Curated desk

the two files that actually interact, a one-line statement of the bug, and the clear ask — nothing else

Where it breaks

The kitchen-sink session: one long thread you keep piling onto for hours. It feels efficient — everything's “in there” — but the desk is now mostly clutter, and answers get vaguer the longer it runs. When a thread starts drifting, start a fresh one with a clean desk rather than fighting the mess.

Try it yourself

Next time an answer is off, don't rephrase — audit the desk. List what the model can actually see right now. The missing piece is almost always sitting in your head, never having made it onto the desk; and the noise is almost always something you could clear off it. Adjust what's on the desk before you touch a single word of the prompt.

Grounded in Anthropic's writing on context engineering — the smallest set of high-signal tokens, and the model's limited attention budget.

Select any text above and sign in to keep a private note on this chapter.

Tap to appreciate

Was this chapter helpful?