The five parts of a prompt that works

Everyone who uses generative artificial intelligence has lived the same scene. You ask for something, receive a text that is technically correct and completely useless, and conclude that the tool does not fit your work.

Most of the time the problem is not the model. It is the instruction. Generative models complete patterns: they produce the most likely continuation of whatever you wrote. Generic instruction, generic continuation. That is why two professionals can use exactly the same tool, on the same day, and get results of completely different quality.

The five parts of a request

Define where the tool should answer from. "You are a senior commercial contracts analyst" produces a different result from no definition at all. The role calibrates vocabulary, depth and what the tool will treat as relevant.

This is the most skipped and the most valuable part. What document is this? Who is the result for? What is the objective behind the task? Context separates an encyclopedia answer from an answer that applies to your work.

Replace generic verbs with operational ones. "Analyse this contract" is vague. "Extract every termination clause and indicate the page where each one appears" is executable and verifiable.

Say how you want to receive it. Table or running text? How many items? Which columns? What length? If you do not define it, the tool decides, and it usually decides on the format that takes the most work to use.

Agree in advance on what will be checked, by whom and against what. Every factual item, whether a number, a date, a clause or a reference, has to be validated at the source. This step is not bureaucracy. It is what makes the use professionally defensible.

The difference in practice

Improvised request:

Result: a generic summary, with no focus, that only becomes useful after you read the whole contract anyway.

Structured request:

Same document. Same tool. Results that cannot be compared.

And there is a detail that changes everything: the second request is reusable. It becomes a template anyone on the team applies to the next contract, with the same output standard and the same review criteria.

What to do after the first result

Work in rounds

The first answer is a draft, not a deliverable. Refine it: "go deeper on item 3", "too long, cut it in half", "rewrite in a more sober tone".

Show the standard you want

Pasting an example of your own material and asking the tool to follow that structure works better than any abstract description of tone.

Ask for the path, not only the conclusion

"Explain which passage of the document supports each item" lets you audit the result in seconds.

Keep what worked

A tested prompt saved in a personal file helps one person. The same prompt documented and shared becomes an asset of the organization.

Three cautions that cannot be ignored

AI is confidently wrong

It can produce numbers, dates, references and citations that do not exist, with a flawless appearance. No prompting technique removes this. Human verification at the source is not optional.

Confidential information requires judgment

Client data, confidential documents and personal information should not be entered into open tools without prior assessment. Check the platform's data handling policy, prefer contracted corporate environments and anonymise whenever possible.

Responsibility remains human

The professional signs the work. AI accelerates production, but it does not transfer authorship, judgment or accountability.

From isolated request to process

Writing good requests for AI is, at bottom, the same competence as delegating well: explain the context, define the objective, agree on the format and establish how the result will be checked.

The difference is scale. When that structure stops being individual talent and becomes a documented standard, with prompts calibrated on the organization's real material, defined workflows and review with criteria, the organization stops depending on whoever has a knack for AI and starts having a process.

Start small. Pick a task you repeat every week, rewrite the request using the five parts, and keep what worked.