Debugging prompts

When the agent gets it wrong, here’s how to diagnose why and re-prompt cleanly.

When a prompt produces the wrong result, the temptation is to argue with the implementation. Don’t. Roll back; refine the prompt; resubmit.

Diagnosis checklist

1. Did the agent restate the prompt correctly?

   Plan mode shows the agent’s restatement. If it misses the point, the prompt was the issue.

2. Did it use the right entity names?

   “the data” instead of “the tasks table” leaves room for misinterpretation.

3. Did it touch only what you asked?

   Over-editing is a sign of missing constraints (“don’t change X”).

4. Did it pick the right integration?

   “send an email” without specifying Resend may pick SendGrid; be explicit.

Common fixes

Add a constraint

“Same change, but don’t break the existing X.”

Switch to Plan mode

The plan reveals the misread before code is written.

Switch agents

Codex for tight changes; Gemini for layout; Claude for everything else.

Reduce scope

Break a 3-thing prompt into 3 separate prompts.

Tips

Roll back before re-prompting. Don’t iterate on top of broken output; you’ll compound the issue.

Related

Iterating effectively

The mindset.

Common mistakes

Patterns to avoid.

Plan mode

The cheapest insurance.