How to Write a Prompt That Works the First Time

Most people iterate three or four times to get usable output. Most of that iteration is supplying information that could have been in the first prompt.

Five components. In this order.

1. The task, stated as a verb

"Write", "summarize", "compare", "critique", "rewrite". Not "help me with" — that produces a conversation about your request instead of the thing you wanted.

One task per prompt. Two tasks produce two mediocre answers.

2. The audience

Who reads this, and what do they already know? "For a technical audience" and "for someone who has never used AI" produce entirely different, equally correct outputs.

Skipping this is the most common cause of output that is technically fine and unusable.

3. The format and length

"Six bullet points." "Under 200 words." "A table with three columns." "An email with a subject line."

Without this, you get the model's default format, which is usually longer and more structured than you wanted.

4. Your material

The source document. The past example. The data. The style sample. This is the single biggest lever on quality, and the one people skip because pasting feels like effort.

Output quality tracks input specificity almost linearly. A model with your material outperforms a better model without it.

5. The constraints

What to avoid. Words you never use. Claims not to make. Things to leave out.

Negative constraints are underused and disproportionately effective — "do not include a summary paragraph" saves an edit every single time.

Putting it together

"Write a 200-word announcement for existing customers, plain tone, based on the release notes below, ending with one clear action. Do not use marketing superlatives and do not add a closing summary."

That is all five components in two sentences, and it usually works on the first try. Everything else — role-playing, elaborate formatting, magic phrases — matters far less than these five.