In a business AI system, the prompt, meaning the set of instructions given to the model, works like a job description. It tells the model who it is, what it must do, with what information, within which limits and in what form to deliver its work. A vague prompt produces vague results, whatever the model.
Yet prompts are often written in a hurry, patched as complaints come in, and never tested systematically. This article presents the method we use to write reliable prompts, and the mistakes we see most often.
A prompt is a brief
The best way to approach a prompt is to write it as a brief for a capable colleague who knows nothing about your company. They are smart and learn fast, but they do not guess. Anything that is not written down does not exist for them.
This simple change of mindset already solves many problems: you stop writing “answer the customer” and start writing what a new colleague would need to know to answer well.
The elements of a good prompt
- Role and context: who the assistant is, for which company, facing which audience. “You help the customer service agents of a medical equipment distributor answer questions about delivery times.”
- A clearly stated task: one main task, described precisely. If the prompt asks for five different things, it is often better to split it into several steps.
- Constraints: what to do and what never to do. Tone, length, excluded topics, information that must never be invented (prices, lead times, contractual commitments).
- Examples: two or three examples of inputs and expected answers are often worth more than a long paragraph of explanation. They should be varied, so the model does not just copy a single case.
- Output format: a precise structure when the answer feeds another tool (named fields, a strict data format), a clear layout when it is meant for a person.
- Permission to say “I don’t know”: state explicitly what to do when information is missing. It is one of the simplest ways to reduce invented answers.
A model does not guess what you expect. Anything that is not written in the prompt does not exist for it.
The most common mistakes
- Vague instructions: “be professional”, “keep it short”. Professional according to whom? Short how? It is better to describe the expected result, or give an example.
- Contradictory instructions: a prompt that has grown through successive patches often ends up asking for one thing and its opposite. The model then decides in its own, unpredictable way.
- Piling up special cases: every complaint adds an “unless…” line. After a few months, nobody knows what the prompt does anymore. A frequent special case often deserves its own step or rule.
- Capitals and threats: writing “VERY IMPORTANT” everywhere does not make an instruction better followed; explaining why it exists is usually more effective.
- Asking the model to do what code does better: calculations, dates, consistency checks. The prompt is not the right place for these.
- Mixing instructions and data: the documents, emails or texts to process should be clearly delimited, and presented as information to analyze, not as instructions. This also protects against prompt injection.
Treat prompts like code
In a production system, a prompt is a component in its own right: changing a few words can change the behavior of the whole system. It therefore deserves the same care as code.
- Versioning: each prompt is stored in the code repository, with a history of changes and the reason for each one. You always know which version runs in production, and you can roll back.
- A test set: a set of representative cases with expected results, including hard cases and trick cases. Every prompt change is validated by rerunning this test set.
- Review: a prompt change is reviewed, just like a code change. A fresh pair of eyes quickly spots ambiguities.
- Production monitoring: cases where a person corrected the result are collected and added to the test set.
A concrete example
A sales team uses an assistant to summarize meeting notes and extract next steps. The initial prompt is one line: “Summarize these notes and list the actions.” Results are inconsistent: summaries too long, invented actions, wrong owners.
The reworked version specifies the context (a B2B sales team), the expected summary length, an output format with one field per action (description, owner, due date), the instruction to leave a field empty rather than guess, and two complete examples. The meeting notes are placed in a clearly delimited section. The prompt is versioned, and a set of around fifty real, anonymized meeting notes is used to validate every change. Invented actions become rare, and when a salesperson corrects a result, the case joins the test set.
Our recommendation
Write your prompts as precise briefs, not magic formulas. Give context, a clear task, limits, examples and an output format, and explicitly plan for missing information. Then treat them like code: versioned, tested, reviewed. It is less spectacular than the tricks that circulate online, but it is what makes the difference between a successful demo and a system that is reliable day after day.