I'm beginning to approach weak AI outputs this way.

When a response fails, I don't simply blame the model. I inspect the instruction.

Maybe the objective lacks clarity. Maybe the context doesn't provide enough information. Maybe the output format creates confusion. Maybe two instructions conflict.

I identify the weak point, change the relevant instruction, and test again.

This process gives me a more structured way to improve prompts.

Prompt → Output → Diagnose → Modify → Retest.

That's becoming one of my practical approaches to prompt optimization.

How do you troubleshoot prompts that don't work?