What to leave out

Every token you send draws on the same attention budget. So the real skill isn't stuffing the window, it's choosing. Anthropic's engineers put the goal as the smallest set of high-signal tokens that gets you the result you want. Smallest, not shortest: if Claude needs a fact, leaving it out is its own failure.

A quick test for each piece of context: would the answer change if this were gone? If not, it's costing you attention and money for nothing.

What usually earns its place: the task, the facts the task depends on, the constraints, the format you want, and one or two good examples.

What usually doesn't. Backstory nobody asked for: how the project started, who sits where. Stale versions: the 2019 rules and the 2021 rules that replaced them. Repeats: the same instruction three times in capitals. A laundry list of edge cases: Anthropic recommends a few varied, typical examples instead, because a model learns more from good examples than from a long list of rules. Whole files when a section will do, and raw logs or tool dumps the model doesn't need yet.

That last one has a better answer than deleting: keep a pointer (a file path, a doc title, a link) and let the model fetch the content only when it needs it. That's how agents like Claude Code work, and it's where Give your model tools and Retrieval without tears pick up.

The left reply picked up the noise: a work anniversary, a broken coffee machine, a policy that no longer applies. Nothing in it was wrong to include on purpose. It was wrong to include by accident.

Your turn. The draft is a real-feeling mess. Cut it down to what the reply needs: the customer's message, what to offer, the tone and a length. The checks can only see what's there, not what you removed, so read your version back before you run it.