A Better Workflow for Iterating on AI Prompts

Most people improve prompts by guessing: add more detail, change the role, ask for a better answer, try again. That can work, but it is inefficient.

A better workflow is to diagnose why the output failed, then change the smallest part of the prompt that addresses that failure.

Step 1: Name the Failure

After a bad answer, classify the problem:

• Missing context
• Wrong audience
• Wrong format
• Too generic
• Too long
• Too confident
• Missing tradeoffs
• Unsupported claims
• Weak examples
• Unsafe recommendation

Do not rewrite the whole prompt until you know the failure.

Step 2: Add One Constraint

If the output is too broad, add a constraint:

Make this specific to solo consultants selling fixed-scope services.

If it is too long:

Keep the answer under 180 words and use three short sections.

If it is too risky:

Flag assumptions and avoid legal, medical, or financial advice.

Step 3: Add an Example

Examples clarify taste faster than instructions.

The output should feel like this example: [paste example].
It should not feel like this: [paste bad example].
Explain the difference before rewriting.

Step 4: Ask for a Self-Review

After the model answers, ask:

Review your answer against the goal, constraints, and audience. What should be improved before I use it?

This often catches tone mismatches and missing pieces.

Step 5: Save the Working Version

When a prompt works, save:

• The prompt
• The input that made it work
• The output format
• The follow-up that improved it
• The failure it solved

This turns prompting from trial and error into a reusable workflow.

Iteration Rule

Change one thing at a time. If you change role, context, format, examples, and constraints all at once, you will not know what improved the output.

Iteration Log Template

Use a short log when a prompt matters:

Prompt version:
Task:
Input used:
What failed:
Change made:
Result:
Keep / revise / discard:

This is useful for teams because it prevents people from repeating the same failed prompt changes.

Example Iteration

Problem: an email prompt keeps sounding too apologetic.

Do not rewrite the entire prompt. Add one constraint:

Keep the tone warm but firm. Do not over-apologize. Acknowledge the request once, then move to the boundary and next step.

If that works, save the constraint. If it does not, add a good and bad example so the model can see the difference.

When to Stop Iterating

Stop when the output is useful enough for human editing. Prompt iteration should reduce work, not become the work. If you need more than three or four rounds, the task probably needs better source material, clearer constraints, or a different prompt type.