How to write prompts that return a usable result: what the model makers advise

A prompt is the whole request, not a magic phrase

Most people who look up how to write prompts expect a list of phrases. The guides of the companies that build the models contain no such list. Wikipedia defines a prompt as natural language text that describes and prescribes the task an AI should perform. OpenAI’s documentation adds the word that matters: instructions written so that the model consistently generates content that meets your requirements. The skill sits in the whole request, and nothing in it is secret.

Why does the same question bring a strong answer on Monday and a weak one on Tuesday? OpenAI states that what a model generates is non-deterministic. Google’s Gemini guide names the mechanism: a temperature setting controls the degree of randomness in the choice of each next piece of text. Wikipedia adds the writer’s share: research shows that models are highly sensitive to subtle variations in the formatting and structure of a prompt, so a gap in the request is filled differently each time.

Five parts of a request: task, context, limits, form, a test of done

Anthropic’s guide offers the most useful picture: a brilliant but new employee who lacks context on your norms and workflows. Its golden rule: show your prompt to a colleague who knows little about the task, and if they would be confused, the model will be too. So the first two parts are the task and the context: who the text is for and what came before. Google’s guide shows the effect on a faulty router: with its troubleshooting guide pasted in, the answer becomes about that device.

The third part is the limits, the fourth the form of the answer. Google’s guide says that you can tell the model what to do and what not to do; its example is one line, summarise this text in one sentence. You can also ask for a table, a list or a paragraph. Anthropic adds two refinements. Say what you want instead of what you do not want. And give the reason: a ban on ellipses works better once the model knows that a speech engine will read the answer aloud.

The fifth part is the one beginners skip: how the result will be judged. Anthropic’s overview assumes that before any tuning you have a clear definition of success and a way to test against it. OpenAI’s advice for its reasoning models agrees: be very specific about your end goal. Write the test into the request: the length, the reader, the facts that must appear.

A worked example inside the request beats a description of it

Describing a tone is hard; showing it takes one paste. Anthropic calls examples one of the most reliable ways to steer format, tone and structure, and recommends three to five. They should be close to the real case, varied and clearly separated from the instructions. Google goes further and recommends always including examples, warning that prompts without them are likely to be less effective. A shop owner can paste two product descriptions that already sell and ask for a third in the same manner.

The makers also state the limits. Google warns that with too many examples the model may start to overfit its response to them. OpenAI asks for a diverse range of inputs, and for its reasoning models suggests trying without examples first. One fact from Wikipedia saves disappointment: what a model learns from examples inside a prompt is temporary, unlike training. In a fresh conversation the lesson is gone, a good reason to keep your examples in a file.

One request, one job: a long task as a chain of steps

A request that asks for research, a plan, a text and a translation at once usually returns four mediocre things. Google’s guide gives the remedy in a sentence: instead of having many instructions in one prompt, create one prompt per instruction. For sequential work it describes a chain, in which the output of one prompt becomes the input of the next. A commercial offer then becomes four steps: the client’s situation, the structure, the draft, a final comparison with the brief.

Anthropic adds a caveat. Its current models, the guide says, handle most multistep reasoning internally, so a chain is no longer needed to make the model think. It stays useful when you want to inspect the intermediate results or hold the work to a fixed order. The most common pattern there is self-correction: a draft, a review against your criteria, a refined version. The gain is simple: a mistake caught at step two does not travel into steps three and four.

Reasoning, questions back and permission to say “I do not know”

The oldest trick is asking the model to think step by step. Wikipedia describes chain-of-thought prompting as a technique that lets a model solve a problem as a series of intermediate steps before the final answer. The makers now say this is often built in. OpenAI’s guidance for reasoning models calls such prompts unnecessary, and Google notes that its recent Gemini models reason internally. Anthropic observes that a general instruction to think thoroughly often produces better reasoning than a hand-written list of steps.

What still earns a line is the check at the end. Anthropic suggests appending an instruction to verify the answer against your criteria, and says this catches errors reliably, especially in code and mathematics. It also notes that one of its recent models does this unprompted, which shows how quickly such advice ages. Its page on reducing invented answers begins with a simpler step: explicitly give the model permission to admit uncertainty. A line such as “if the report lacks the necessary information, say so” changes what comes back.

Questions back work the same way. A new employee would ask before starting, and OpenAI describes its reasoning models as ones that will often ask clarifying questions before making uneducated guesses. With any other model one line requests this: before you write, ask me what you need to know. Afterwards, ask what the model was unsure about and what it assumed. None of this makes the answer true; it shows where to look first when you check it.

Correcting a draft instead of starting the conversation over

Google’s guide states it as a principle: prompt engineering is iterative. The first answer is a draft. When it misses, do not wipe the conversation; say what is wrong, what should stay and what the next version must contain. If that fails, Google lists three moves: use different words, switch to an analogous task, such as turning an open classification into a multiple-choice question, or change the order of the content in the prompt.

A request that finally works is worth more than the answer it produced. Save it with the example and the test that went with it. OpenAI advises teams to build tests that measure how a prompt behaves, to monitor performance as they iterate or when the model version changes. The one-person version is a note beside each saved prompt: the date, the model, one input and the output you accepted.

What a prompt cannot fix: missing facts, the wrong tool, an unchecked answer

Anthropic’s documentation is direct about the limit. Even the most advanced language models can sometimes generate text that is factually incorrect or inconsistent with the material they were given. Its techniques reduce such errors without eliminating them, and the closing note is blunt: always validate critical information, especially for high-stakes decisions. Wording cannot supply a fact the model does not have. OpenAI’s guide describes the remedy as adding context: your own documents and data from outside the model’s training.

Some tasks need a different tool, not a better sentence. Google’s guide says that grounding with search should be enabled whenever the model may need obscure or recent facts, and code execution whenever it has to do arithmetic, counting or calculation. A bare chat model is the wrong place for this week’s exchange rate or a column of sums. Anthropic adds that not every failing result is best solved by prompting. On law, tax or health, the answer is material for a specialist to check, never the decision itself.

Your own prompt library and what a structured course adds

This is enough for a week of practice alone. Pick one task you repeat: replies to reviews, a product description, a weekly report. Write the request in five parts, add two examples of a result you would accept, and run it on three real cases. Correct it, then save the version that held up. A dozen such entries make a working library with your tone, your products and your tests. Whoever still asks how to write prompts after that week is really asking where to put them to work.

That next question is what a course is for. The Artificial Intelligence course of VITON13 SCHOOL teaches the deliberate use of AI, from framing the task through prompting and verification to automating your own processes and building agents. Its modules: Working with AI deliberately, Prompting as engineering, Research and verification, Automating processes, AI agents, Your own systems, Safety and responsibility. It is an online course with practice on the learner’s own task. Self-study gives you a library; a programme puts the topics in order. The course page carries the current terms.