You don't write good prompts. You revise them

1. The reframe

Nobody writes a perfect prompt on the first try. Not seniors, not the people who write prompt-engineering articles, not the model's own builders. The skill isn't writing. The skill is revising.

A first prompt is a hypothesis. The output is the experiment result. The second prompt is the hypothesis adjusted by what you learned. Most people stop after the first turn — either accept the imperfect output or abandon the model. The people who get useful work out of AI are running two to four iterations on every non-trivial task.

2. What revising looks like

Revision isn't try again with feeling. It's a deliberate diagnostic — the same skill you'd use debugging code.

A good revision:

• Reads what came back specifically — not wrong but this part is wrong because...
• Identifies which part of the original spec the model didn't honor.
• Adds the missing constraint or context to the prompt — doesn't restart.
• Leaves the parts that worked alone.

A bad revision:

• Throws out everything and starts over with the same shape of prompt.
• Adds vague pressure: try again, really try this time, be more careful.
• Doesn't read the first output carefully — just notices the vibe is off and reruns.

3. Revise vs. restart

The first prompt asks for a function. The output has a correct signature but wrong error handling.

That's wrong. Try again.

The model rolls dice. Often the next attempt has the same problem, or a new problem, because the model doesn't know what was right about the previous attempt and what was wrong.

The signature and the happy-path logic are correct — keep both.

The error handling is wrong: when 'input' is null, we should return null (not throw). When 'input' is an empty array, we should return [] (not throw). The current code throws in both cases.

Fix only the error handling. Don't change anything else.

The model now has a specific diff to make and explicit instructions on what to preserve. The second attempt is closer, not random.

4. The reframe in practice

When you send your first prompt, send it knowing it's the first of three (probably). The first sets up the structure. The second corrects what the model missed. The third tightens or extends.

This reframe alone — I'm sending the first of three — changes how you write prompt one. You include less because you know you'll add later. You read the output more carefully because you know you'll act on it. You stay engaged instead of disappointed.

5. Try it