Prompt Iteration: Fixing Prompts That Don't Work

Prompt design is rarely perfect on the first try. You write a prompt, check the response, and adjust. That is completely normal, even experts do it. Here are three strategies to fix prompts that are not giving you what you want.

Strategy 1: Rephrase Your Prompt

Different phrasing often gives different results, even when the meaning is the same. If one version does not work, try saying it differently:

• Version 1: "How do I bake a pie?"
  Version 2: "Suggest a recipe for a pie."
  Version 3: "What is a good pie recipe?"

All three ask for the same thing, but the model may respond differently to each. Sometimes the second or third version gives a noticeably better answer.

Strategy 2: Switch to an Analogous Task

If the model will not follow your instructions one way, try framing the same goal as a different type of task. For example, if you ask the model to categorize a book into predefined genres and it gives you a long explanation instead of just the category, reframe it as a multiple choice question.

Instead of: "Which category does The Odyssey belong to: thriller, sci-fi, mythology, biography?"

Try: "Multiple choice problem: Which option describes The Odyssey? Options: thriller, sci-fi, mythology, biography."

Same goal, different framing, and often a cleaner answer.

Strategy 3: Change the Order of Content

The order of things in your prompt can affect the response. Try rearranging the sections:

• Version A: Examples, then context, then your question
  Version B: Your question first, then examples, then context
  Version C: Examples, then your question, then context

Bonus tip: If the model gives you a fallback response like "I'm not able to help with that," try increasing the temperature slightly. This can sometimes get past overly cautious safety filters.

Based on: Gemini API Prompting Strategies (ai.google.dev)