How to Write a Prompt That Works on the First Try

You ask for a summary of a client call. You get three paragraphs that restate the transcript: no decision, no owner, no date. You add "be more specific." You get four paragraphs. You give up and write the note yourself.

The model was not the problem. The brief was. Most advice on how to write a better prompt skips the obvious: a prompt is a brief, not a wish. A prompt that works on the first try is not a magic sentence to stumble onto. It is a short, complete instruction built from the same parts you would give a capable new hire: who they are, what you need, what they should know, what to avoid, what shape the answer takes, and one example. Fill those in and the first draft usually lands close enough to edit instead of redo. Not clever phrasing. Complete context.

Why the first draft missed

Look at the gap between what you typed and what you pictured. Usually it is one of three things.

You pictured a reader the model did not know about. "Summarize this call" for your own notes is a different job from a summary for a VP who has 90 seconds and needs to make a decision.

You pictured a shape. You wanted a decision line, owners, dates. The model wrote prose because you never asked for anything else.

You pictured judgment. You knew the renewal was sensitive, that the client was already unhappy, that the last summary caused a round of confusion. None of that was in the prompt, so none of it shaped the answer.

None of this needs a better model. It needs a better brief.

How to write a better prompt: the six part brief

The six part brief is a fixed order you can fill in for almost any task. It takes about two minutes once you have done it a few times.

• Role. Name the perspective you want the answer from. "You are an operations lead writing internal notes" beats "you are a helpful assistant," which tells the model nothing. The role sets word choice, what gets noticed, and how much explanation is assumed.
• Task. One verb, one object, one purpose. "Turn this transcript into a decision brief so my VP can approve or hold the renewal." If you cannot say it in one sentence, you have two prompts.
• Context. Everything the model cannot infer: who reads the output, what decision it feeds, what happened before, what already failed. Paste the source material instead of describing it. A pasted transcript or email is worth more than a paragraph summarizing it.
• Constraints. Length, tone, what to leave out, and what to do when information is missing. That last one matters more than people expect. "If no owner was named, write unassigned. Do not guess" prevents the most annoying kind of wrong answer.
• Output format. The exact shape, section by section, in order. If you want a heading called Decision, say so. If you want bullets under each heading, say that too. A skeleton of the output is easier to follow than adjectives.
• Example. One short sample of the answer you want, or a before and after. This is the highest leverage part and the one people skip. An example settles tone and level of detail in a way that no amount of describing can.

You do not need all six every time. A quick rewrite needs task and format. Anything that feeds a decision, goes to another person, or touches a customer deserves all six.

A weak prompt, rewritten

Here is the kind of prompt people send.

Summarize this sales call and make it good.copy

And here is the same request as a brief.

You are an operations lead who writes tight internal notes.

Task: turn the transcript below into a decision brief for my Monday leadership meeting.

Context: we are deciding whether to renew the fulfillment contract. The reader is my VP, who was not on the call and has 90 seconds. The client raised two complaints about late shipments.

Constraints:
- 200 words maximum.
- Only include commitments someone actually made.
- If no decision was reached, say that in one line.
- Do not invent owners, dates or numbers. Write unassigned or unknown instead.

Output format:
- Decision: one sentence
- Commitments: one bullet per item, with owner and date
- Open questions: bullets
- Next step: one sentence

Example of the tone I want:
"Decision: hold the renewal until the SLA report lands. Owner: Priya, Friday."

Transcript:
[paste]copy

The second version is longer, but length is not the point. Each line closes a door the model would otherwise walk through. The weak version leaves the reader, the decision, the shape and the escape hatch all undefined, so the model fills the gaps with averages. That is why the output reads like a summary of a summary.

The prompt engineering template I reuse

This is the skeleton I copy for anything that matters.

Role: You are [specific role] writing for [reader].

Task: [one verb] [one object] so that [reader] can [decision or action].

Context: [what happened]. [what the reader already knows]. [what has been tried or ruled out].
The source material is below.

Constraints:
- Length: [number of words or bullets]
- Tone: [plain, formal, blunt, warm]
- Exclude: [anything that should stay out]
- If information is missing: [say unknown / ask me / leave the field blank]

Output format:
[section name]: [what goes here, in what length]
[section name]: [what goes here]
[section name]: [what goes here]

Example of the output I want:
[one short sample]

Source:
[paste]copy

Fill it in once for a task you repeat, like call summaries or customer replies, and you have a prompt you can trust without rewriting it every Monday. If you want a filled version with more examples, the free prompt kit has them ready to copy, and the Prompt Engineering Field Manual goes deeper on the same six parts across writing, research and analysis work.

Two habits make the template hold up. Keep the example in, even after the prompt starts working; that is usually when quality drifts. And version the prompt like a document: change it deliberately and keep the old version until the new one survives a week.

Check it before you hit enter

A pre-send checklist catches most first draft failures in under a minute.

• Can I name the reader without thinking? If not, the model cannot either.
• Is there exactly one task? "Summarize and critique" is two prompts wearing one coat.
• Did I paste the source material, or did I describe it? Paste it.
• Does the format name every section, in order, with lengths?
• Did I say what to do when a fact is missing? If not, expect a confident guess.
• Is there one example of the tone I want?
• Would a smart new colleague have to ask me a question to do this? If yes, answer that question in the prompt.
• Read it out loud. If it sounds like a job posting, cut the adjectives.

The checklist is about noticing the gaps before the model fills them for you.

Failure modes that waste the most time

Most bad first drafts trace back to a handful of habits.

Two tasks in one prompt. Summarize and then critique, or draft and then fact check. The model splits its attention and does both halfway. Split them into two prompts and run them in order.

Tone by adjective. "Make it professional" means nothing on its own. Professional to a law firm and professional to a startup are different documents. Show one sentence of the tone you want instead.

Format described in prose. "Give me a table" when the output lands in a chat window, a doc or a spreadsheet gets you three different results. Name the medium and the sections.

No escape hatch. If the prompt does not allow "unknown," the model will produce something plausible. Plausible is worse than blank when the output feeds a decision.

Context without labels. When you paste a long email thread, say which part is the instruction and which part is reference. Otherwise the model treats a customer's complaint as a request aimed at you. If you routinely paste text from the web, email or customers, read this on defending an agent against injected instructions before you wire the prompt into anything that takes action.

Fixing by adding "be more concise." Repeated tone corrections rarely work. The real miss is usually the format or the example. Fix the part, not the adjective.

When the output still misses

Sometimes the brief is good and the answer is still off. Before you rewrite everything, check which part was ignored. If the sections are wrong, the format was too vague. If the tone is wrong, the example was missing or too generic. If facts appear that you never supplied, the escape hatch was missing.

Change one part, run it again, and compare. Rewriting the whole prompt hides which change helped, and you end up with a prompt that works by accident. Keep the version that survives two different inputs. One good result on one document is a coincidence; a prompt that holds on three is a tool.

Questions people ask

Do I need a different template for each model?

No. The six parts are about the task, not the vendor. Models differ in how they follow long instructions and render structure, so check the format on the model you actually use. The brief stays the same.

How long should a prompt be?

As long as it takes to remove the ambiguity, no longer. A rewrite of one paragraph can be three lines; a brief that feeds a hiring decision may run a page. If you keep adding words and the output does not improve, you are adding noise.

Does a good prompt mean I can skip checking the output?

No. A good prompt reduces how often you have to redo the work. It does not verify facts, and it does not make a wrong source right. Read the output the way you would read a draft from a competent colleague: fast, but not blind.