Skip to main content
The message box is not the prompt. It is the last line of the prompt. When you press send, Claude receives one long document: its standing instructions, your project’s instructions, any files in view, every earlier message in the thread, and then, at the very bottom, the sentence you typed. It continues from all of it, weighted more or less evenly. There is no separate channel for “background.” Everything in view is instruction.
SystemInstructionsFilesChat history
what you typed

One prompt, to scale. The sentence you wrote is the sliver on the right. The answer is a continuation of the whole bar.

If the answer is a continuation of the whole bar, then changing the bar changes the answer more than changing the sliver.

​
Same message, three desks

Here is the identical request sent three times: “Write the follow-up email to the client.” Only the desk changes. Switch desks, then click any item on the desk to see exactly which parts of the answer it produced.
Claude did not get smarter or dumber between desks. It did the same thing all three times: continue from what was in view.
  • The empty desk forces guessing. Every specific in that answer is a placeholder wearing a suit.
  • The cluttered desk is worse than empty. Wrong context does not get ignored, it gets used. Claude has no way to know the deck and the old thread are irrelevant.
  • The briefed desk produces the answer that sounds like you knew what you wanted. Because, in the prompt Claude actually read, you did.

​
The habit

Before sending anything that matters, look at the desk instead of polishing the sentence:
  1. What does Claude need in view to do this well?
  2. What is currently in view that it should not use?
  3. Where will it get the standard for what “good” looks like?
A mediocre question on a briefed desk beats a brilliant question on an empty one.