How Much Context Is Enough? Give the Model What Changes the Answer

More context is not automatically better. The useful question is: which facts would change a correct answer? If I ask AI to review a bug, the error message and the code path matter. The color of my editor usually does not. For sensitive tasks, pair useful context with Privacy Checks Before You Paste Anything Into AI so you do not paste data the model never needed.

Suppose a game score disappears after a page refresh. This is too little context: Why does my score disappear? This is more useful: In a browser game, score is held only in a JavaScript variable. After refresh it returns to 0. Explain why that happens, compare localStorage with a server-side save for this use case, and recommend a first implementation. Do not assume we already have an account system. Now the model can explain that in-memory state does not survive a reload and compare persistence options. But I should not paste every file in the project. The prompt already describes the behavior and the missing decision.

Take each sentence in your context and ask: “If this were different, might the recommendation change?” An account system would change the storage decision, so it matters. The exact shade of the score button would not. Keep context that changes the answer; drop the rest. For code-specific questions, attach the smallest complete snippet that reproduces the problem, the exact error if there is one, and the expected versus actual result. For factual questions, identify the source document and date. For creative questions, provide the audience and constraints that shape the work.

Do not guess at ten irrelevant details. Ask: Before recommending a fix, ask up to three questions whose answers would materially change the implementation. If no such question is needed, state your assumptions and proceed. That keeps the exchange focused. Sometimes the best context is an answer to a clarifying question. Try it: Take one prompt you use regularly. Cut its context in half. Keep only facts that pass the counterfactual test, then compare both answers against the same rubric. If the shorter version still passes, save it. Keep learning: Break a Large Request Into Smaller Prompts, Prompt AI to Explain Code Without Inventing What It Does, and Privacy Checks Before You Paste Anything Into AI. Further reading: clear-instruction guidance.