Prompt Fundamentals: Goal, Context, Constraints and Format

Suppose you hire a freelancer and send a one-line message: "Write something about our new product." You will get a reasonable attempt, and it will be wrong in ways you did not anticipate. Wrong length, wrong tone, wrong audience, wrong angle. Nobody is surprised, because you gave the freelancer almost nothing to work with.

AI assistants behave the same way. The quality of the output depends heavily on what you gave it. This article teaches what to give.

• Why vague prompts produce generic answers.
• The four elements of a clear prompt, and a fifth for when quality matters.
• How to improve a weak prompt step by step.
• What to make of "act as an expert" and other folklore.
• A template you can reuse tomorrow.

When you leave something unstated, the model does not leave it blank. It fills it with the most typical choice. Ask for "an email to my landlord" and you get a typical landlord email: generic opening, no real details, medium formality. That is not laziness. It is the best guess about what you probably want when it knows nothing about you.

So the whole skill can be summarised in one sentence: replace the model's average guesses with your actual information.

Every serious provider says some version of this. OpenAI's prompt guidance emphasises giving clear instructions and relevant context. Google's guidance asks for clear, specific instructions, response formats and constraints. GitHub's advice for Copilot is to start with a broad description of the goal, then list specific requirements. They are all describing the same craft.

And a fifth element for tasks that matter:

You rarely need all five for a casual question. For anything you will actually use, aim for at least goal, context and format.

Problem. The water heater in your rented flat has not worked for a week. You want to message your landlord.

First attempt.

Write an email to my landlord about the water heater.

Example response (illustrative).

Subject: Water Heater IssueDear Landlord,I hope this email finds you well. I am writing to inform you that the waterheater in my apartment is not working. Please let me know when it can berepaired. Thank you for your attention to this matter.Sincerely,[Your Name]

What is missing. No dates, no history, no urgency, no request for a specific action, no tone choice, and it says "apartment" when you say "flat". It is also generic enough to be ignored. The assistant did nothing wrong. It had nothing else to go on.

Improved request. Now add the four elements:

Goal: write a message to my landlord asking for the water heater to berepaired this week.Context: I rent a flat in Pune. The heater stopped working 7 days ago. Ireported it by phone on Monday and he said he would send someone, but nobodycame. We have a good relationship and I want to keep it. The rentalagreement has a clause that says repairs under Rs 3,000 are hisresponsibility.Constraints: polite but firm; under 120 words; do not threaten or mentionlegal action; do not invent any dates or amounts beyond what I wrote.Format: a WhatsApp message, plain text, with a one-line ask at the end.

Example response (illustrative).

Hi Mr. Kulkarni, the water heater in the flat stopped working 7 days ago. Icalled on Monday and you kindly said someone would come, but no one has yet.Since hot water is a daily need, could you arrange a technician by Friday? Asper our agreement, repairs like this are covered on your side, and I'm happyto share photos or be at home whenever they visit. Thanks for your help.

What changed and why. Every addition replaced a guess. The context gave real facts. The constraints set the tone and blocked invented details. The format made the output ready to send. You still need to read it: the name "Mr. Kulkarni" was a placeholder you should have supplied or removed, and you should check that the sentence about the agreement matches your actual clause. Example responses are illustrations. Yours will differ, and none of it is guaranteed to be correct.

A language model continues text in the most plausible way given everything in front of it. A short prompt matches thousands of different situations, so the most plausible continuation is a bland average. A detailed prompt matches far fewer situations, so the plausible continuation is much closer to what you actually want. Detail is not ceremony. It narrows the model's search.

This also tells you what not to do. Adding words that carry no information ("please give me a really great, amazing, detailed answer") does not narrow anything. Add information, not length.

The landlord message was a writing task. Here is the same method on an explanation task, where the failure looks different.

Problem. You are preparing for an interview and need to explain how a hash map works.

First attempt.

Explain hash maps.

You get a generic textbook paragraph. It may be accurate, and it does not match what you need: it is too long, assumes the wrong level and gives you no way to tell whether you understood it.

Improved request.

Goal: help me explain hash maps clearly in a job interview.Context: I am a final-year student. I know arrays and loops in Java, but I do not knowwhat a hash function is. I have 20 minutes to prepare.Constraints: at most 250 words; one everyday analogy; mention one downside; no code.Format: three short paragraphs (what it is, how a lookup works, one trade-off), thenask me one question to test whether I understood.

Example response (illustrative). Three short paragraphs using a library-shelf analogy, a mention that two keys can land in the same slot (a "collision"), and a closing question such as "What happens if two keys land in the same slot?"

What to check. The analogy is a teaching aid, not a technical fact, so ask where it breaks down. The claims about speed, for example "constant time on average", should be checked against your course notes or the documentation of the language you use, since the details depend on the implementation. The closing question is the most valuable part: answer it without looking, then ask for feedback.

The four elements are always present, but one of them usually matters most, depending on the type of task:

When an answer disappoints, name the symptom and look for the matching gap before you rewrite everything:

Fixing the specific gap is usually faster than starting again, and it teaches you which of your prompts tend to have which gap.

You can use this skeleton for most tasks. Delete the lines you do not need.

Goal: <what I want and what it is for>Context: <who is involved, background, what I have already tried>Constraints: <length, tone, things to include or avoid, what to do if unsure>Format: <table, bullets, steps, email, JSON, etc.>Check: <how I will judge a good answer, or ask it to list assumptions>Material: <paste the text, notes or data it should work from>

You will see advice such as "Act as a senior marketing consultant." Does it work? In a limited way. Anthropic's prompt engineering overview, for example, lists role prompting among its techniques, and giving a role can shift tone, vocabulary and focus. "Explain this as a patient teacher speaking to a 12-year-old" changes the register in a useful way.

What a role does not do is add knowledge or make reasoning more reliable. "Act as a doctor" does not give the model medical training or make its answer safe. A better habit is to describe the audience and the criteria:

A role is a style instruction. Treat it that way.

When the task is ambiguous, you can avoid a bad first answer by inviting questions.

Before you write anything, ask me up to three questions that would most improveyour answer. Then wait for my replies.

This is useful when you do not know what information is relevant, which is common in unfamiliar areas. It turns the model into an interviewer for the first turn. Keep the limit ("up to three"), or you may get a long list.

Prompts that ask for five different things tend to do each one worse, and you cannot easily see which part went wrong. Compare:

• "Summarise this report, list the risks, rewrite the conclusion, and draft an email to the team."
• Four separate requests, each checked before the next.

You can still chain them, and article 9 shows how. The point here is to keep one clear job per request when quality matters.

• "Say please / thank you for better results." Politeness is fine for you; do not count on it to change quality.
• "Add 'take a deep breath' or offer a tip." Any effect, if there is one, is likely to be small and specific to one model and one moment. Do not build habits on tricks like these; build them on information.
• "Longer prompts are always better." Relevant detail helps. Padding does not.
• "There is a perfect prompt for every task." There is a prompt that works well enough for your situation, and you improve it by looking at what went wrong.
• "Provider X's prompt tricks work everywhere." Models differ, and so does advice. Check the documentation for the tool you use.

• Starting with the format and forgetting the goal. "Give me bullet points" is not a goal.
• Hiding the real purpose. If the answer is for a hostile audience or a strict deadline, say so.
• Leaving out what to do when unsure. Add "if you do not know, say so" or "ask me".
• Pasting everything you have instead of the relevant part. More text is not more context.
• Accepting the first answer as final. It is a first draft. Article 9 covers refining.
• Forgetting to check. A perfectly formatted answer can still contain invented details, as in the landlord example.

Goal: suggest 10 video ideas for my channel that could help it grow from 200 to1,000 subscribers over the next 3 months.Context: the channel teaches Excel to people starting office jobs in India. Ihave 12 videos so far; the two that did best were "VLOOKUP in 5 minutes" and"Excel shortcuts for beginners". I film on my phone, alone, and can make onevideo a week.Constraints: each idea must be doable in under 4 hours of work; nointerviews or travel; avoid topics that need paid software.Format: a table with title, the first sentence I would say, and an effortrating (low, medium, high).Check: after the table, tell me which idea you are least sure about and why.

Audience: recruiters for entry-level software roles in India who read a resumefor about 30 seconds.Goal: tell me which three parts of my resume are most likely to be skipped ormisunderstood, and why.Criteria: a good answer points to specific lines in my text, not general advice.Constraint: do not add any experience or skills I did not write.[paste resume]

• Next: Give AI the Right Context: Source Material and Long Documents shows how to supply the material the model should work from.
• To practise on your own tasks, start your playbook with the template in this article.
• For a shorter companion guide, see Writing Better AI Prompts: A Practical Guide.