Know What Actually Fits in the Context Window

A working understanding of the context window: the finite budget of tokens Claude can attend to at once, shared across the system prompt, everything you've pasted, the conversation history, and the space left for the reply. It is short-term working memory, not permanent storage.

People treat the context window as either infinite ('paste everything') or as memory that persists between chats — both wrong, and both cause failures. Knowing it's a shared, finite budget explains why a huge paste degrades answers (signal gets diluted, and the reply gets starved of room), why Claude 'forgets' something from far back in a long chat, and why a fresh chat doesn't remember the last one. That model tells you when to trim, when to summarise, and when to start over.

Any long conversation, large-document work, or task where you're tempted to paste a lot. Understanding the budget is what makes context engineering possible.

Short, self-contained prompts where you're nowhere near the limit — you don't need to think about the budget for a two-line question.

I'm working with <describe: a long doc / a long chat / a big codebase>. Before I paste, help me budget the context window: what's essential to include, what I can summarise instead, and what I should leave out — so the answer has room and the signal isn't diluted.

Instead of pasting a 200-page manual to ask three questions, you paste the two relevant chapters; the answer is sharper because Claude isn't hunting for your questions inside a haystack, and there's plenty of window left for a full response.

For work that outgrows the window, keep a running summary you carry forward — distil the conversation to its decisions and open questions, start fresh with that summary as the seed, and you get continuity without dragging the entire history along.

• Assuming a bigger window means you should fill it — dilution still hurts quality.
• Expecting a new chat to remember an old one; context doesn't persist unless you re-supply it.
• Pasting so much input there's little room left for a thorough answer.