A good prompt is a brief, not a wish. It says what problem you are solving, what you want back (the format, the length and who will read it), the limits the answer has to stay within, the assumptions it should state rather than make silently, and what counts as done. When one of those is missing, AI usually fills the gap instead of asking. It fills it with whatever a request like yours usually gets, which is why a weak prompt so often comes back sounding fine and helping nobody.

Why a vague prompt still gets a confident answer

An AI model is built to produce a likely continuation of what you gave it. Ask for “an update on last quarter” and it writes the update a typical company might send: smooth, upbeat and generic, because nothing in your request tied it to your company. The draft is not badly written. It is empty, and that is harder to spot than an obvious error.

Anthropic’s own guidance for Claude puts it this way: treat the model like a brilliant but new employee who lacks context on your norms and workflows, and be specific about the output format and constraints you want. Its suggested test is a good one for any assistant: show the prompt to a colleague who knows nothing about the task. If they would be confused, the AI will be too.

The five parts of a brief

PartThe question it answersWhat goes wrong without it
Frame the problemWhat is this for, and what decision or action follows from it?You get a general essay on the topic instead of help with your situation.
Say what you want backWhat format, how long, and for which reader?It picks its own structure, length and tone, pitched at an average reader who does not exist.
Set the limitsWhat must it use, and what must it not do or promise?It adds things nobody approved, or draws on general knowledge where you needed your own figures.
Expose assumptionsWhat should it say it is unsure of, instead of guessing?Gaps get filled with plausible guesses that read exactly like facts.
Define doneHow will you know the answer is ready to use?You cannot tell a finished answer from a nearly finished one, so revising never ends.

Before any of the five, make sure you are asking for one job. “Help me sort out the supplier situation” is a week of work with a title on it. One job is one thing to make, for one reader or group, for a reason, with a clear point where AI’s part stops and yours starts.

A weak prompt, rewritten

Here is a realistic example. You run operations at a sample office-supply distributor. Your finance director, Dana, has asked why you want to switch your main packaging supplier, and she will decide on Thursday whether to approve the new contract. The wish version of the prompt looks like this:

Write something about why we should switch packaging suppliers.

That will come back as a tidy list of reasons any company might switch suppliers. Now the brief:

I run operations at a mid-sized office-supply distributor. Our finance director, Dana, will decide on Thursday whether to approve a contract with a new packaging supplier. She needs to know why I am recommending the switch.

Write a one-page memo for Dana. Start with a two-sentence recommendation, then three short sections: cost, delivery reliability and risks. Plain language, under 400 words.

Use only the figures in the comparison table below; do not estimate any others. Do not suggest a switch date, because none has been agreed. Leave staffing out: it is not part of this decision.

If the table is missing something you need, say what is missing rather than filling it in. List any assumptions you made at the end.

The memo is done when Dana could approve or reject the contract from this page alone, and every number in it appears in the table.

[comparison table pasted here]

Each paragraph is one part of the brief. The first frames the problem and names the decision. The second sets the format, length and reader. The third sets the limits, including two things the AI could easily have invented (a date and a staffing angle). The fourth tells it to expose gaps instead of covering them. The last says what done means in a way you can check.

Tip: Say why a limit exists, not only what it is. “Do not suggest a date, because none has been agreed” tells the AI the reason, so it also avoids nearby mistakes, such as “we expect to switch next month.”

Fix the brief, not only the answer

When the answer misses, the tempting move is to patch it in the chat: “shorter,” “less formal,” “add the risks.” That can work for a one-off, but each patch fixes this answer and teaches you nothing about the next one. A better habit is to ask which part of the brief was missing, add it to the prompt and run it again.

Fixing the brief also pays off when you repeat the job. A brief that works once can be saved and reused, and if you do the job every week it is worth turning into a skill your assistant can reuse (see what a Claude skill is).

A quick checklist before you send

None of this needs special phrasing. It is the same brief you would give a capable new colleague on their first day, written down.