Curate the Context — Don't Just Dump It

Context engineering: deliberately choosing what Claude sees — the relevant files, facts, and constraints — rather than pasting everything and hoping the model finds the signal.

A model reasons from what's in front of it. Padding the context with loosely-related material lowers the signal-to-noise ratio and lets the model anchor on the wrong thing. Curated context is why the same question gets a sharp answer one time and a muddled one the next.

High-stakes or ambiguous tasks, long documents, and codebases — anywhere the difference between a good and a bad answer is which facts the model had, not how you phrased the ask.

Small, self-evident prompts where there's barely any context to curate.

Here is exactly the context you need for this task — nothing more, and I've left out what's irrelevant on purpose:

<curated facts / files / constraints>

Task: <ask>. If something you'd need is missing, ask for it rather than guessing.

Instead of pasting a whole repo, you give Claude the three files that own the bug and the failing test — and get a precise fix rather than a plausible edit to the wrong module.

Ask Claude what additional context would most change its answer before it commits. Letting the model name its own missing inputs is a fast way to find the gap in what you provided.

• Equating more context with better answers and drowning the signal.
• Leaving in outdated or contradictory material the model then treats as true.
• Never asking what's missing, so the model fills the gap with a confident guess.