Fixing a bad output without starting over

When an answer is wrong, most people retype the whole prompt. Don't. The conversation already contains useful context — steer instead of restarting.

The four steering moves

• "Too long — cut it to 80 words, keep the statistics." (correction + constraint)
• "You assumed I sell B2B. I sell to hobbyists. Rewrite for them." (correct the context)
• "Give me three versions: one formal, one friendly, one blunt." (diverge, then pick)
• "What information would make this answer better?" (flip the interview)

When to restart instead

If the model misunderstood the task itself (not the details), a clean start with a sharper prompt wins. Patching a misunderstood task piles patches on patches.