Good AI Prompt: The Anatomy of an Effective Prompt in 5 Blocks
A good AI prompt has five blocks: a role, context, a precise task, constraints and an expected output format. When one is missing, the model fills the gap with whatever it considers most likely, which usually means a generic answer. This article breaks down each block with before/after examples so your teams get usable answers on the first try. It complements our complete prompt engineering guide for SMEs.
Why a vague prompt gets a vague answer
A language model cannot read your mind: it continues the text you give it. An LLM predicts the most plausible continuation of your request. If the request is one line long, the most plausible continuation is an average answer, valid for everyone and therefore useful to no one.
Take this request: "Write a customer follow-up email." You will get a polite eight-line text with no name, no amount and no date. Three quarters of it will need rewriting. The time saved is cancelled by the time spent fixing it, which is why many employees conclude that AI "isn't that useful". In most cases the problem is the instruction, not the tool.
A well-built prompt also lowers the risk of hallucination: the more you frame the request, the less the model needs to invent to fill the blanks.
Block 1: the role, to set register and expertise level
The role tells the model who it is supposed to be, which sets vocabulary, level of detail and tone. One sentence is enough: "You are the office manager of a 25-person company" does not produce the same answer as "You are a corporate lawyer".
This is the shortest block and the one most often forgotten. It does not make the model more competent, but it steers it toward the right register. For team use, define three or four standard roles (sales, HR, accounting, support) and reuse them.
One limit to keep in mind: a role never replaces verification. Asking AI to "act as a lawyer" gives its output no legal value.
Block 2: the context, what the model cannot guess
Context gathers the information specific to your situation: who the recipient is, what happened before, what the goal is. It is the block that makes the biggest difference, because the model knows nothing about your company, your customers or your last exchange.
Good context answers four questions:
- Who reads the result (a customer, a colleague, a funder)?
- What happened before (first contact, second reminder, dispute)?
- What concrete outcome do you expect (a payment, a meeting, a decision)?
- Which data must the model use (amounts, dates, names)?
Be careful about what you paste into the chat window. Customer data, contracts, health information: not everything can be sent to a consumer tool. Our article on data confidentiality with AI details the rules to set, and the one on shadow AI explains what happens when those rules do not exist.
Block 3: the task, one clear action verb
The task describes what the model must produce, using a precise verb: write, compare, summarize, classify, extract, rephrase. Avoid instructions that pile five requests into one sentence: the model usually handles two or three well and quietly neglects the rest.
Two simple habits improve results:
- One main task per prompt. If you need a summary and then an action plan, send two messages.
- An explicit success criterion: "The text must let the customer pay without calling us back".
For complex cases, break the work into steps. This is the principle of guided reasoning, which we will cover in a dedicated article in this cluster.
Block 4: the constraints, to avoid off-target answers
Constraints set the limits: length, tone, forbidden words, mandatory elements, allowed sources. They rule out what you do not want, which context alone does not guarantee.
Useful constraints for small businesses:
- "Maximum 120 words, formal tone, no excessive apologies."
- "Do not cite any figure I did not give you. If data is missing, write [TO COMPLETE]."
- "Do not mention our competitors."
The second constraint deserves to be made a habit: explicitly asking the model to flag what it does not know is one of the most effective ways to limit made-up content. It is not a guarantee, but it noticeably reduces the risk.
Block 5: the output format, for a ready-to-use result
The format states the exact shape of the deliverable: table, numbered list, email with subject line, JSON, three options to compare. Without it you receive a block of text that someone has to reformat by hand.
Ask for the format you will actually paste somewhere. A four-column table for a spreadsheet, an email with subject and body separated for your mail client, JSON for an automation tool. In that last case rigor matters: one misnamed field is enough to break a scenario.
Before / after: a reminder email in 5 blocks
The gap between a weak prompt and a structured one shows immediately on a real case.
Weak version: "Write a reminder email for an unpaid invoice."
Structured version:
Role: you are an administrative assistant in a 20-person services company. Context: the customer Dupont SARL has a 2,400 EUR invoice overdue by 18 days. This is the first reminder. The relationship is good and the customer usually pays within 30 days. Task: write a reminder email that makes them want to pay without offending them. Constraints: formal tone, 100 words maximum, no blame, mention the invoice number [TO COMPLETE]. Format: an email subject line, then the message body.
The first version produces a text you rewrite. The second produces a draft you review in thirty seconds. The figures in this example are illustrative: adapt them to your own files.
A simple quality test: give the prompt to a colleague who does not know the case. If they lack information to answer, so does the model.
The most common mistakes in companies
The same mistakes recur from team to team, and most can be fixed in minutes.
| Mistake | Consequence | Fix |
|---|---|---|
| Request too short | Generic answer | Add context and recipient |
| Several tasks in one | Parts ignored | One prompt per task |
| No format requested | Manual reformatting | Specify table, list, subject + body |
| Sensitive data pasted | Confidentiality risk | Anonymize or use an approved tool |
| Answer accepted without review | Errors get published | Systematic human review |
| Prompt never reused | Everyone reinvents it | Shared prompt library |
The last point is a team lever. When a prompt works, it should be saved, named and shared. That is what turns individual use into a company practice.
From a good prompt to a team skill
Writing a good prompt can be learned in half a day, but making it a habit takes a framework. Employees progress fast when they work on their own files rather than generic examples.
It is also a compliance topic. Article 4 of the EU AI regulation (Regulation (EU) 2024/1689) requires organizations that use AI systems to ensure a sufficient level of AI literacy among their staff. Our article on the Article 4 training obligation covers its scope.
To go deeper by tool, see our content on ChatGPT training for business and Claude training. If your organization is small, AI training for micro-businesses offers a suitable format. For the overall framework, see the complete guide to AI training for business.
How GrowthPerf supports your teams
GrowthPerf trains your teams to write and share prompts on their own business cases. We are a Qualiopi-certified training body, which may open the possibility of OPCO funding depending on your situation (eligibility to be confirmed case by case).
In practice, a session starts from three tasks your teams do every week, builds the matching prompts with them, tests them, then stores them in a shared library. Our prompt engineering training details the program.
To know where to start, book a free 30-minute AI training audit: we review your current uses and tell you what deserves framing first. Then find the whole approach in the prompt engineering guide.
Frequently asked questions
How long does it take to write a good prompt? Count two to five minutes the first time, then almost nothing once the prompt is saved and reused. The investment pays off on repetitive tasks.
Do you always need all five blocks? No. For a simple question (rephrase a sentence), task and format are enough. The five blocks are for requests where quality matters.
Is a long prompt better than a short one? Not by nature. A long, confused prompt performs worse than a short, precise one. What matters is that every sentence adds useful information.
Does the same prompt work on ChatGPT, Claude and Gemini? The structure often does, but results vary from one model to another. Test your prompt on the tool you actually use.
How do you stop AI from making things up? Provide the source data, forbid figures you did not supply, ask it to flag what is missing and always review. No instruction removes the risk on its own.
Can you put customer data in a prompt? It depends on the tool and your GDPR framework. Anonymize whenever possible and set a clear internal rule before any rollout.
Does a good prompt replace training? It is one building block. Training brings practice, a security framework and collective adoption, which reading an article alone does not.