Why can't I get reliable results, even with a good prompt?
You treat the model like a vending machine. Put in the right words, get the right answer out. When the answer is wrong, you assume the prompt was wrong, so you throw it out and write a new one from scratch. But you never examine what came back. You don't ask which part failed, or why it failed that way. You just discard and retry, hoping the next attempt lands by luck. That is not a prompting problem. That is a missing habit. The skill you lack has a name: look at the output, find exactly where it diverges from what you needed, change one thing to close that gap, try again. Do this enough times and you stop guessing. You start knowing.
You are not bad at prompting. You are missing a feedback loop. You write, you get something back, and you move on without asking what was wrong or why. Build the habit: look at the output, name the gap, change one thing. That is the entire skill.
What changes unlock by starting
- Within two tries, you can tell whether a bad result is the model's limit or your unclear ask.
- You need fewer attempts to get work you can actually use.
- You build a written record of what kind of instruction this model responds to.
- You stop blaming yourself or the tool, and start adjusting instead.