The quality of an artificial intelligence answer depends largely on the quality of the request. A prompt, the text you use to ask a language model for something, works rather like the instructions you would give a new colleague who is highly skilled but knows nothing about you, your business or why you need the work done. The clearer the picture, the less the model has to guess.
Why vague prompts produce generic answers
If you write "write me an email for a client", the model does not know who you are, who the client is, what happened, what tone to use or how long it should be. Faced with so many unknowns it takes the most probable route, which is the most average one: a polite, generic and nearly useless email. This is not a lack of intelligence; it is a lack of information.
The elements of a good prompt
An effective prompt usually contains five components. You do not always need all of them, but keeping them in mind helps.
- Context. Who you are, who the result is for, what the situation is. "I run a small engineering practice; a client is complaining about a two-week delay in delivering a project."
- Goal. What you really want to achieve. "I want to apologise, explain the cause and propose a new date without sounding defensive."
- Format. Length, structure, tone. "No more than 150 words, professional and friendly tone, no bullet points."
- Material. Data, texts or documents to work on. Paste the client's email, your notes, the table.
- Constraints. What to avoid. "Do not promise discounts, do not mention the supplier."
A practical comparison
| Weak prompt | Effective prompt |
|---|---|
| "Summarise this document." | "Summarise the minutes below in 5 points for a manager who did not attend the meeting. Highlight decisions made, owners and deadlines." |
| "Fix the code." | "This Python function should return the sum of the even values, but with the list [2, 3, 4] it returns 9. Find the bug and explain the fix." |
| "Give me marketing ideas." | "I own a neighbourhood hardware shop whose customers are mostly over 50. Suggest 5 low-cost initiatives for the autumn, with an estimate of the effort each requires." |
Techniques that work
Give examples
If you want a particular style or format, show one. Two or three examples of product descriptions you have already approved teach the model more than any explanation of "brand tone". This technique is called few-shot prompting.
Ask it to reason step by step
For multi-step problems, such as a cost calculation or the analysis of a contract, asking the model to proceed one point at a time and show intermediate steps makes any errors easier to spot.
Assign a role, in moderation
Phrases such as "act as a careful reviewer" can steer style and level of detail. A role does not, however, add knowledge the model lacks: it helps with tone, not with accuracy.
Separate instructions from material
When you paste a long text, mark it off clearly, for example between triple quotes or after a line reading "TEXT TO ANALYSE:". That way the model can tell what it has to do from what it has to process.
Iterate
The first result is rarely the final one. Instead of starting over, ask for targeted changes: "shorten the second paragraph", "make the closing more formal", "add a numerical example".
Common mistakes
- Piling up requests. Ten different tasks in a single message lead to shallow answers. Better to split them.
- Contradictory instructions. "Be thorough but extremely brief" forces the model into an arbitrary compromise.
- Assuming shared context. The model does not know what you said in another chat or what is in your files, unless the service explicitly connects them.
- Trusting the result without checking. An excellent prompt reduces errors; it does not eliminate them.
- Including confidential data without knowing how the service handles it.
A reusable prompt template
You can adapt this structure to many tasks:
- Context: "I am [role] and I am working on [situation]."
- Task: "I need you to [precise action]."
- Material: "Here is the data: [text or table]."
- Format: "Reply with [structure, length, tone]."
- Constraints: "Do not [things to avoid]. If information is missing, ask me before proceeding."
That last sentence is particularly useful: inviting the model to ask questions before answering brings out what you forgot to say.
Prompts and language
The most widely used models also work well in languages other than English. If you are working on technical texts in one language, it can help to write the prompt in the same language as the material and ask for the answer in the language you need. Keep terminology consistent: if you call something a "job order", do not call it an "order" later in the same prompt.
In short
- A good prompt provides context, goal, format, material and constraints.
- Concrete examples and step-by-step reasoning greatly improve answer quality.
- Mark off the material to be analysed and work through successive iterations.
- Avoid piled-up requests, contradictory instructions and unprotected confidential data.
- Even the best prompt requires a final check of the result.
