The Prompt Failed. What Should You Change First?

The fastest way to improve a prompt is to name the failure. “The model was bad” does not tell me what to change. “It summarized the wrong document section” does. Once you make one change, Compare Two Prompt Versions Without Fooling Yourself helps you test whether that change actually improved the result.

Wrong job: You asked for “help” when you needed a review, comparison, or draft. Fix the task verb and define what should be included. Missing decisive context: The model assumed a user account, a date, or a game mechanic you never supplied. Add the one fact that changes the answer or ask it to clarify. Wrong output: The answer is a long essay when you needed a checklist or code with tests. Specify the form that makes the result usable. No acceptance test: You cannot tell whether the answer is good. Write a criterion before revising the prompt. Suppose the model gives you generic game ideas when you want a mechanic you can prototype tonight. The likely failure is not its writing style. The task lacks a prototype constraint. Try: Propose three mechanics for a two-player game that can be tested with paper cards in one hour. For each, show the player action and the first 5-minute test. Reject any mechanic that requires new art, networking, or a long story. Now an idea is useful only if it passes the buildability check.

If you simultaneously change the task, audience, format, model, and context, you may get a better answer but learn nothing about why. Change the most likely cause first. Keep the original answer beside the new one. Google’s prompt design strategies frame prompting as experimentation and refinement rather than a fixed incantation.

If the model lacks access to a live fact, a current document, or your actual code, adding clever wording will not supply it. Bring the needed evidence, use an appropriate tool, or do the check yourself. Try it: Save one disappointing answer. Label the failure as job, context, output, test, or missing evidence. Make one targeted edit and compare against the original criterion. Keep learning: Make a Bad Prompt Better in Three Edits, How to Test a Prompt on More Than One Example, and Compare Two Prompt Versions Without Fooling Yourself.