Put the Most Important Context Where Claude Looks

A placement habit for long prompts: put the material that most determines the answer at the start or the very end of the context, and let bulk reference material sit in the middle. Order is not cosmetic — where a fact sits changes how strongly it lands.

Attention across a long context is uneven. The beginning sets the frame and the end is freshest when Claude starts writing; the middle of a large block gets the least weight — the well-documented 'lost in the middle' effect. If the deciding constraint is buried on page 7 of 12, it competes with everything around it. Moving it to just before your instruction, or leading with it, makes Claude weight it the way you intend. This is free leverage: same tokens, better placement.

Any prompt with more than a page or two of context: long documents, multiple files, a big spec plus a small ask. Especially when one constraint or one source should dominate the answer and the rest is background.

Short prompts where everything fits in a glance — position is irrelevant when there's barely any context to lose things in. Don't reorder for its own sake on small inputs.

<Bulk reference material — the documents, code, or data Claude may need.>

---
The single most important constraint for this task: <the deciding fact or rule>.

Now, using the material above and giving that constraint priority, <your specific request>. Format the answer as <format>.

You paste a 4,000-word policy plus a short email to answer. Buried in the middle, the clause 'refunds are void after 30 days' gets missed and Claude approves a refund. Move that clause to a line directly above your question — 'Given that refunds are void after 30 days, is this request valid?' — and Claude answers correctly, because the deciding rule is now where it looks.

For repeated long-context work, adopt a fixed template: reference block first, a short 'what matters most here' summary just before the task, and the instruction and output format last. A stable structure means Claude always finds the signal in the same place, and you stop re-discovering placement per prompt.

• Leading with a wall of reference material and tacking the actual question on at the end with no signposting — the key constraint drowns.
• Assuming Claude weights every sentence equally because it 'read' the whole context; capacity to read is not the same as attention to act on.
• Repeating the important fact five times to force it through, instead of placing it once in a high-attention slot.