How to Write a Prompt That Works on the First Try

People blame the model. Usually the request was the problem. You asked for "a summary" and got a summary — just not the one in your head, because the one in your head had a length, an audience, and a purpose that never made it into the message. Running that same vague request through a prompt enhancer button does not fix this either — it dresses the request in more words without adding the missing facts.

Here is what a prompt that lands actually contains. None of it is clever. All of it is specific — and it matches what both OpenAI and Anthropic tell you in their own prompting guidance. It is also most of what a paid prompt engineering certification teaches in its first module.

1. The job, stated as a job

"Help me with this email" is not a job. "Rewrite this email so it fits in four sentences and does not sound annoyed" is. The difference is that the second one can be checked. If you cannot tell whether the output succeeded, the request was too vague to succeed.

A good test: could a competent stranger do this task from your description alone, without asking you a follow-up question? If not, the model has the same problem. It knows nothing about your company or your reader unless you tell it, for reasons that come straight out of what GPT stands for.

2. Who it is for

Audience decides vocabulary, length, and how much context to assume. "Explain our pricing change" produces something entirely different when the reader is an existing customer versus your own finance team. Say which.

3. The shape of the answer

Ask for the format you actually want: three bullets, a table with these columns, one paragraph, a numbered checklist. If you do not specify, you get the model's default, which tends to be longer and more hedged than anything you would send. That padded house style is one of the failure modes in what AI is actually bad at.

• —Length — "under 120 words", not "short".
• —Structure — bullets, table, prose, code block.
• —Tone — "plain and direct", "warm but brief".
• —What to leave out — "no preamble", "do not restate the question".

4. An example, if you have one

This is the highest-leverage thing in the list and the one people skip. Paste a version you liked — an old email, a paragraph from a doc, a competitor's page — and say "match this". One example carries more information about tone — the behaviour these models were built around — than a paragraph of adjectives ever will.

The same four things matter even more once the task grows from one answer into a multi-step research request — deep research prompts covers what changes when you are asking a model to plan its own queries and synthesise across many sources rather than answer in one pass.

Describing a style takes three sentences and lands at seventy percent. Showing one takes three seconds and lands closer to ninety.

The habit that matters more than any of this

When an answer is wrong, do not argue with it. Start over.

A long back-and-forth where you keep correcting a bad first draft tends to produce a worse result than one clean second attempt. Arguing also invites the model to simply agree with you, which is not the same as being right, because every message you send is still attached to the flawed one. The model is working from a conversation that is mostly a record of getting it wrong.

So: read the bad output, work out which of the four things above you left out, put it in, and ask again from scratch. That single habit fixes more answers than any phrasing trick. Bear in mind that a confident wrong answer reads exactly like a right one. If you are not sure whether the output is wrong in the first place, how to check an AI answer when you are not the expert covers that.

A worked example

Weak: "Summarise this report."

Better: "Summarise this report for our head of sales, who has not read it and has two minutes. Five bullets, under 100 words total. Lead with anything that changes what her team should do this quarter. Skip methodology."

Same model, same report. The second one is usable without editing, and it took twenty seconds longer to write. The same brief works on newer assistants too, as the worked prompts in this Muse Spark tutorial show. That is why best AI tools for business says the cheapest improvement is rarely a new product. ChatGPT prompts for business applies the same four-part shape to five specific office tasks, if a worked template is more useful than the general rule. A cheat sheet for Excel shows the same weak-versus-better pattern applied to a single SUMIF formula, where the fix is naming the exact columns instead of describing the task in general terms.

That is also most of what sits behind best AI prompts — five reusable ones, each with a real input, a usable result, and the check to run before trusting what comes back, built on the same four-part rule above rather than a new one.

The same four things apply when the output is a picture rather than a paragraph. How to write a ChatGPT photo prompt that actually works walks through the image version with a before/after example, and AI image prompts: a framework that works across tools covers the reusable six-part structure behind it. Even a prompt built purely for a laugh runs on the same rule — a specific constraint beats a vague plea, whether the plea is "be professional" or "be funny".

Coursium teaches exactly this — including practice lessons where you take a weak prompt of your own and rewrite it. Stay ahead of AI by learning the tools on your phone.