Why is a good prompt usually the second or third one?

You will rarely write the perfect prompt on the first try, and that is normal. Prompting is an iterative process. The pattern is simple: send a prompt, read what comes back, then change the wording, add context, or simplify the request based on what you saw.

This loop is built into how the professionals work too. OpenAI’s guide treats prompts as something you test and refine, recommending representative fixtures, tests, and evaluation checks before changing a prompt that runs in production.^[1] The point holds even for a single chat: the first answer tells you what your prompt actually communicated, which is often not quite what you meant.

A reliable way to improve a draft is to have the model help. Anthropic describes a self-correction pattern: generate a draft, review it against your criteria, then refine it based on that review.^[2] In a chat, that just means replying with “good, but make the tone more formal and cut it to two paragraphs” instead of starting over.

When you iterate, change one thing at a time when you can. If you rewrite the whole prompt at once and the result improves, you will not know which change did it. Small, named adjustments teach you what this model responds to, and that knowledge carries to your next prompt.

There is a choice buried in this loop that is easy to miss: refine in the same conversation, or start over in a new one. Refining in place is usually right, since the model can see exactly what it got wrong and adjust from there instead of guessing at your intent again from scratch. But once a session has piled up several failed attempts and corrections, that history is now part of what the model reads on every turn, and a long trail of “no, not like that either” can crowd out the instructions that actually matter. When refinement stalls instead of converging after two or three tries, a fresh conversation with one carefully written prompt often beats a sixth correction to a conversation that has drifted far from where it started.