Prompting That Works

Most people iterate three or four times before getting something usable. Most of that iteration is supplying information that could have been in the first message. Prompting well is mainly about front-loading.

Forget magic phrases and elaborate role-play. Five components account for nearly all the difference.

1. The task, as a verb

Start with what you want done: write, summarize, compare, critique, rewrite, extract, classify. Not 'help me with' — that invites a conversation about your request rather than the thing itself.

One task per prompt. Asking for a summary and a translation and a social post produces three mediocre outputs instead of one good one.

2. The audience

Who is this for, and what do they already know? 'For a technical audience' and 'for a shop owner with no technical background' produce completely different — and equally correct — answers to the same question.

Leaving this out is the most common cause of output that is technically fine and practically unusable.

3. 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 defaults, which tend to be longer and more elaborately structured than you wanted.

4. Your material

The source document. A past example in your voice. The actual data. The real customer message. This is the biggest single lever on quality, and the one most people skip because pasting feels like effort.

A weaker model with your material will beat a stronger model without it. Almost every time.

5. Constraints — especially negative ones

What to avoid, what not to claim, what to leave out. Negative constraints are underused and disproportionately effective, because they remove the specific habits that make output feel generated.

Write a 200-word announcement for our existing customers, in a plain tone, based on the release notes below, ending with one clear action.

Do not use marketing superlatives, do not add a closing summary paragraph, and do not claim anything the release notes do not state.

Release notes:
[paste them here]

Iterate, do not restart

When an answer is close but wrong, the instinct is to rewrite the prompt and start again. That throws away the context you just built.

Follow-up instructions compound: 'shorter', 'less formal', 'keep point two and cut the rest', 'now add a concrete example'. Each one operates on what is already there. Restarting means re-explaining everything.

Outline before expanding

For anything longer than a few paragraphs, ask for the structure first. Adjust it. Then expand section by section.

This catches direction problems when they cost you ten seconds, rather than after the model has written two thousand words in the wrong direction.