Treat every first output as a diagnosis. Read the defect, change one thing, and keep the diff.
Nobody writes the best version of a prompt first. Professionals iterate, and they iterate in a specific way: read the output as a diagnosis, change one variable at a time, and keep the version that worked. The failure mode is not iterating too little; it is iterating vaguely, rewriting the whole prompt every time and never learning which change helped.
Read the output as evidence
When a result disappoints, name the defect precisely. Too long, too vague, wrong audience, invented detail, missing a field, buried the answer. A named defect points at a missing constraint or an unclear task. "It's not good" points nowhere. The discipline of naming a defect is most of the skill, because a named defect can be fixed with one sentence and a vague complaint leads to a rewrite that fixes nothing.
Change one thing
Each revision should change one thing: add a constraint, move a block, replace an example, tighten the task verb. Then compare. If you change three things and the output improves, you do not know which one to keep, and the next revision is a guess. This is the same reason engineers prefer small diffs: a small change has an attributable effect.
Example: an output is too formal for a newsletter. Do not rewrite the whole prompt. Add "Write in a conversational second person, short sentences, no corporate vocabulary." Run it. If it is still too formal, the constraint is not strong enough or the role is pushing the other way — now you have a specific next question.
Keep the version that worked
Save prompts like code. A one-line changelog and a version number cost nothing and make a later revert possible. PromptSewa prompt listings keep an append-only version history for this reason: a prompt is an asset, and an asset without history is an asset you cannot improve safely. When a prompt is used in production, the version that produced accepted output should be identifiable later.
Know when to stop
Iteration has diminishing returns. When the output meets the stated constraints and a reviewer could not name a defect without inventing a new requirement, stop. Endless polishing is how prompts become fragile: each extra clause narrows the model's options until it fails on the next input. Robust prompts are usually a little looser than the most polished version of a single case.
Checklist
- Can you name the defect in one phrase?
- Did the last revision change exactly one thing?
- Is the working version saved with a changelog?
- Are you polishing, or adding a requirement?