Pattern 5 of 10

Respect the context window

A bigger context window tempts you to paste everything. But past a point, more text makes results worse and more expensive, not better.

More text is not better. Keep prompts lean and put key parts first and last.

The context window is everything the model considers at once: your instructions, pasted data, examples, and the running conversation. Respecting it means including what is relevant and cutting what is not.

Why it works

There is also a direct cost angle: longer prompts and longer conversations cost more to process every single time they run. At scale this compounds quickly.

How to do it

  • Include only what the task needs; cut tangential background.
  • Put the most important instructions at the start and end.
  • In long chats, start a fresh conversation when the thread gets bloated.
  • Iterate: start with the shortest prompt that works, add only what fixes a real gap.

Example you can copy

Lean over bloated
Instead of pasting an entire 40-page handbook and asking
"summarize the refund policy", paste only the refund section:

CONTEXT (refund policy only):
<paste the 2 relevant paragraphs>

TASK
Summarize the refund policy in 3 bullet points for a customer.

When it helps

  • Long documents, large codebases, long-running chats
  • Production prompts that run many times, where cost compounds
  • Any time output quality drops as the prompt grows

When to skip it

  • Genuinely needs the full document (then place it well and keep the question at the end)

Common mistakes

  • Pasting everything "just in case", which adds noise and cost
  • Letting one chat run so long the early context crowds out the task
  • Burying the real question in the middle of a huge paste

Frequently asked

Are longer, more detailed prompts always better?

No. Past a point, extra text adds noise, can bury key instructions, and raises cost. Reasoning quality can degrade before the technical limit, so lean, well-placed prompts often beat large ones.

What is the context window?

It is everything the model considers at once: your instructions, pasted data, examples, and the ongoing conversation. Keeping it relevant improves both quality and cost.