How to Write a Good AI Prompt

A good AI prompt tells the model five things: what you are trying to achieve, the background it needs, the exact task, the limits it must respect, and the shape of the answer you want back. Write those down in plain sentences, in that order, and most vague or generic replies disappear. Length matters less than completeness: a prompt is finished when a capable stranger could do the job from it without asking you anything.

Start from the stranger test

Anthropic's documentation suggests treating the model like a brilliant new colleague who has no context on your norms, your audience or your project. That framing is the fastest check available. Read your draft and ask: if I handed this to a smart contractor who has never met me, what would they have to guess?

Every guess is a place where the model will fill in the most average answer it can. "Write a product description for our water bottle" forces guesses about the buyer, the tone, the length, which features matter and where the text will appear. None of those are hard for you to state. They are simply missing.

The five parts of a complete prompt

You do not need labels or special syntax for these, although labels help once a prompt grows past a paragraph.

1. Goal

Say what the output is for. "This is going on a product page read by commuters comparing three bottles" changes the answer more than any adjective you could add to the task line. The goal lets the model make sensible decisions about things you did not think to specify.

2. Context

Give the facts the model cannot know: who the audience is, what already exists, what you have tried, the source material it should work from. If you have a document, notes or data, paste them in rather than describing them. OpenAI's guidance describes this as providing reference text, and notes it reduces the chance of invented details because the model is asked to work from what you supplied.

3. Task

State the action with a verb and a scope. "Summarize" is a verb; "summarize the three decisions made and who owns each" is a task. If the work has stages, list them in the order you want them done.

4. Constraints

Constraints are where you prevent the failure you have already seen. Typical ones:

• Length: a word range or a number of items, not "short".
• Sources: "use only the notes below; if something is missing, say so instead of filling the gap".
• Tone and reading level: name the reader, not just the mood.
• Exclusions: what to leave out, such as pricing, competitor names or a recap of the question.

Where you can, phrase a constraint as what to do. "Write in flowing paragraphs" is easier to follow than "don't use bullet points", a point Anthropic's guidance makes about steering format.

5. Output format

Describe what the reply should look like: headings, a numbered list, a table with named columns, JSON with named keys. If the format is fiddly, show a skeleton of it. A format you describe is one you no longer have to fix by hand.

A worked rewrite

Here is the kind of first draft most people type:

Write a product description for our water bottle.

And the same request with the five parts filled in:

I'm writing the product page for a 750 ml insulated steel bottle sold
to cycle commuters. Shoppers on this page are usually comparing two or
three bottles and skim on a phone.

Facts you can use (do not add others):
- Keeps drinks cold for a working day, hot through a morning commute
- Fits standard bike cages
- Leak-proof lid, one-handed opening
- Dishwasher-safe, 10-year warranty

Write the description for the top of the page.

Constraints:
- 70 to 90 words, second person, plain language
- Lead with the benefit a commuter cares about most
- No claims beyond the facts above; no superlatives

Format: one short paragraph, then three bullets of five to eight
words each.

The second version is longer, but every added line removes a guess. Notice that it also contains no flattery, no "act as a world-class copywriter" and no capital letters. Clear information does the work.

Order and layout

When you include long material such as a transcript or report, vendor guidance generally recommends placing the material first and your instructions or question after it, and separating the two clearly. Delimiters do that job: triple quotes, a labeled block or XML-style tags such as <notes> and </notes>. They stop the model mistaking your source text for instructions and make it easy to refer back to "the notes" by name.

For short prompts none of this is needed. Two or three well-formed sentences can carry all five parts.

Common mistakes

• Asking for quality instead of describing it. "Make it engaging" gives the model nothing to act on. Say what engaging means here: opens with a specific problem, uses one concrete example, no preamble.
• Stacking unrelated jobs. A prompt that asks for research, a draft, a title and a social post usually does each one thinly. Split them, or see the guide to prompt chaining.
• Leaving the reader out. The same explanation written for a specialist and for a newcomer are different documents. Name the reader.
• Hiding the important line. If one requirement matters most, put it on its own line rather than in the middle of a paragraph.
• Treating the first reply as final. A first prompt is a draft. When the answer misses, work out which of the five parts was thin and fix that; the guide on improving AI responses walks through the diagnosis.

A checklist before you send

1. Could a stranger do this without asking me a question?
2. Have I said what the output is for and who reads it?
3. Is the source material pasted in, and clearly separated from my instructions?
4. Is there a number attached to length?
5. Have I said what the reply should look like?
6. Is there anything I would be annoyed to see in the answer that I have not ruled out?

If you would rather fill in fields than write from a blank page, the prompt builder assembles these parts for you, and the library's structured prompts, such as the blog writing prompt or the meeting notes prompt, show the same pattern applied to specific jobs.

How much is too much

Add detail until the stranger test passes, then stop. Extra background that does not affect the answer dilutes the instructions that do, and a constraint you do not care about is one more thing the model may over-apply. When you find yourself writing the same long prompt repeatedly, that is the signal to turn it into a reusable template with blanks, which is covered in prompt templates vs prompts.