Strip the Context Down to What Changes the Answer

A discipline of removing context that doesn't change the answer before you send the prompt. The test for every file, paragraph, or example is simple: if I delete this, would a correct answer change? If not, it's noise — cut it.

Irrelevant context is not neutral. Every extra document is something Claude has to weigh, and unrelated material pulls the answer toward tangents, invites Claude to 'use' details that don't matter, and buries the parts that do. A tight context also leaves more of the window for Claude's actual reasoning and output. Less, but relevant, beats more-but-diluted almost every time — the goal is signal density, not completeness.

Before any context-heavy prompt: pasting code, long threads, research dumps, or several documents at once. Especially when a first answer wandered off-topic — that's usually a signal problem, not a reasoning problem.

When you genuinely don't yet know which parts are relevant — for open-ended exploration or discovery, breadth can be the point. Curate once you know what the task actually hinges on.

I'm about to ask you to <task>. First, help me curate the context.

Here is everything I *could* include: <list the candidate files/sections>. Tell me which of these are actually needed to answer well and which are noise, and why. Then I'll paste only the ones that matter.

Debugging a failing checkout, you're tempted to paste the whole 30-file module. Instead you include only the payment handler, the one failing test, and the error — three files. Claude isolates the bug in one pass, because nothing on screen competes with the signal. The other 27 files would only have offered wrong things to blame.

Have Claude do the pruning: give it the task and a manifest of available context, and ask it to request only what it needs before answering. This turns curation into a cheap first step and often surfaces that you were about to over-supply — or under-supply — the one file that mattered.

• Equating 'gave it everything' with 'gave it enough' — completeness is not relevance, and extra material actively costs you.
• Leaving stale or contradictory context in the window, so Claude reconciles conflicts you didn't mean to introduce.
• Pasting an entire chat history when only the last decision matters — old, superseded turns pull the answer backward.