The prompt guide
Learn to talk to an AI: methods, examples, a quiz and ready-to-copy templates.
Why the prompt matters
A generative AI doesn't “know” what you want: it guesses from what you write. A vague prompt gets a vague answer; a precise prompt gets an answer you can use.
Tell me about home insurance.
Explain the essential cover of home insurance to a first-time tenant. 3 paragraphs max, with an everyday analogy, and end with the most important thing to check before signing.
What turns an average prompt into a good one
Three moves: give context (who you are, what it's for), set the format (list, table, code, length), and scope the task (what you want, and what you don't).
None of this is technical: it's the same skill as writing a good brief or clear instructions for a colleague.
The 5 pillars of a good prompt
Whatever method you use next, these five building blocks always come back. Check them before you send.
Who you are, your situation, your level, what you already have. “I'm new to cooking, I can make pasta and that's it.”
A clear action verb: explain, correct, summarise, generate, compare. One task at a time (or numbered).
Table, list, 3 paragraphs, commented code, JSON… Whatever format you don't ask for, the AI picks.
Length, tone, language, style, what to avoid. “No jargon”, “in English”, “200 words max”.
One example of what you expect is worth ten lines of description — that's the basis of few-shot (page 6).
Method 1: roles and constraints
The simplest and most effective method for everyday tasks: give the AI a role and frame the answer with explicit constraints.
“You are a secondary-school maths teacher used to preparing revision sheets” or “Act as a demanding editor of a blog post”. The role steers the vocabulary, level, tone and type of answer. The same text reviewed by “a teacher” or “a journalist” will not get the same comments.
Everything the answer must respect: length, format, tone, forbidden content, angle to avoid, language. Constraints settle the choices the AI would otherwise make on its own.
Full example
Summarise this economics course.
You are a tutor preparing a student for an economics exam.
Summarise this course in 10 key points, one per line.
Constraints:
- each point must fit in a single sentence;
- technical terms in brackets with a 5-word definition;
- end with the 2 classic exam traps on this chapter.
Method 2: the CREATE method
A mnemonic so you forget nothing: five letters, five questions to ask yourself before writing. Ideal for important or reusable prompts.
What role does the AI play? “You are a kind and methodical fitness coach.”
The task itself, as a verb: “Create a 4-week training plan for a beginner.”
What resembles what you want: “My current schedule: 30 min in the evening, 3 times a week.”
Constraints and corrections: “No gym equipment, in English, warm-up included.”
The expected format: “As a table: columns Week, Goal, Exercises.”
The same need, with and without CREATE
Make me a revision plan for my final exams.
[C] You are a study coach preparing final-year students for their exams.
[R] Draw up my revision plan for the next 6 weeks.
[E] My current schedule: 2 h every evening, plus Wednesday afternoon.
[A] My weak subjects: philosophy (8/20) and maths (10/20); I'm aiming for a high grade.
[T] Answer as a table (Week, Philosophy, Maths, Other), then 3 tips as a list.
Method 3: iterative prompting
You don't write a good prompt on the first try: you build it through dialogue. Every imperfect answer tells you what needs to be clarified.
Follow-ups that work
They aren't questions, they are correction instructions:
“Too long. Keep the essentials in 5 bullet points.”
“Good, but use a more direct tone, without polite filler.”
“Point 3 is wrong: redo it taking [correction] into account.”
“Use exactly the same structure as what you did for section A.”
“Start again.”
“This is bad.”
“Do better.”
(without saying what to fix — the AI will make the same choices again)
For an ambitious result, first ask for an outline, approve it, then ask for the writing section by section. You fix an outline in 10 seconds; you fix a full text in 20 minutes.
Few-shot: show rather than describe
Few-shot prompting means giving one or more examples of the expected result in the prompt. The AI imitates the pattern instead of guessing.
Example: rewriting titles
Rewrite my titles to make them catchier.
Rewrite my titles following this model:
“Easy recipes” → “15 recipes ready in 20 minutes for weeknights”
“Holiday budget” → “A week away for under €500: the guide”
My titles:
“Guitar lessons” → ?
“Gardening for beginners” → ?
A style to imitate (marketing tone, email format, structure of a sheet), classifications (“here are 3 correctly labelled cases, classify the next ones”), conversions (JSON to table…). Two or three examples are almost always enough; beyond five, the effect levels off.
For multi-step problems (maths, diagnosis, planning): “Solve this step by step, showing your reasoning before the final conclusion.” Explicit reasoning greatly reduces logic errors — and lets you check where the AI went wrong if the result is incorrect.
Skills, tools and context: beyond the prompt
A good prompt doesn't stop at the text: modern assistants accept documents and have tools and “skills” — knowing how to call on them is part of the craft.
Rather than describing your project in 50 lines, attach the documents (PDF, spreadsheet, screenshots) and ask the AI to refer to them: “Here is my course (attached). Using only this, prepare 20 flashcards for me.” The AI works on the real data instead of imagining it — that's the difference between an answer about your case and a generic one.
Skills (abilities built into an assistant, such as an in-depth research mode or a specialised workflow) are sometimes triggered just by mentioning them: “do in-depth research on…”, “analyse this data file”, “create a website for…”. Some assistants automatically load the skill suited to your request — you only have to be clear about the intent. The habit to build: when a task is repetitive (the same report format every week), ask the assistant whether a skill exists or how to automate it rather than describing everything again each time.
If the assistant can search the web, open pages, do calculations or handle files: ask for it explicitly. “Look up the current price before answering”, “calculate precisely, don't estimate”, “generate the file and show it to me”. Without instructions, the assistant may answer from memory — unreliable for anything that changes (prices, versions, news, dates).
Tips and pitfalls
The details that change everything: eight habits of a seasoned prompter.
“Don't be too technical” leaves the AI to guess where the limit is. Say what you want instead: “explain with everyday words, as if to a curious friend”. Positive instructions are followed more faithfully than prohibitions.
Don't make it too long. No jargon. Don't start with “Here is”.5 one-line bullet points maximum, in simple words. Start directly with the first point.When you mix instructions and content (text to translate, code to fix), separate them with delimiters: triple quotes, ### or tags. The AI then knows exactly what is instruction and what is data.
Translate this text into French, professional tone:
"""
[your text here]
"""
“Analyse this report, extract the key figures, write a summary for management and suggest 3 actions”: better to ask in 2–3 messages, or as numbered steps in the same prompt. Each sub-task then gets proper attention.
“Don't explain what BPM is again” avoids answers pitched at the wrong level. You can also paste the start of a conversation elsewhere: “for context, I've already done this: …”.
“Suggest 3 versions: a sober one, a punchy one, an original one” — you then choose, and the iteration starts from a better base than a single attempt.
“Here is my prompt: …. What information are you missing to answer well? Ask me 3 questions.” Simple, and remarkably effective for important tasks.
If your prompt contains a factual error (“summarise article 12 of decree XYZ” when it doesn't exist), the AI may go along with it instead of pointing it out. For verifiable topics: “check what I'm stating and point out any inaccuracy”.
“Be creative”, “stay factual”, “be concise” really change the style of the answer. Choose it deliberately rather than putting up with the default style.
Test yourself!
Five situations, five choices. Click the best version of the prompt.
Score: 0 / 5
The final checklist
Before sending an important prompt, go through it. Tick as you review.
Ready-to-copy templates
Want ready-made prompts? The prompt builder offers 62 ready-to-use prompts to fill in, and a guided builder with a quality score.
Skeletons to adapt: replace the brackets, keep the structure. Each template applies the methods from the previous pages.
You are [role: private tutor / coach / expert colleague].
Explain [concept] to me for a [level: complete beginner / intermediate / expert].
Context: [why you're learning it, what you'll do with it].
Format: [3 paragraphs / 10 bullet points / analogy + example].
Constraints: simple words, one concrete example, end with [what I must remember / a common pitfall].
Here is my document: [attachment or text between """..."""].
Summarise it for [use: revising on the train / presenting to my manager / quoting in a dissertation].
Format: [N one-line bullet points / 200 words / table: key idea, detail, page].
Keep exact definitions in brackets. Ignore [parts to leave out].
Write [type of text] for [audience].
Goal: [inform / persuade / sell].
Tone: [direct and informal / professional / enthusiastic].
Length: [N words max].
Structure: hook, [N] arguments with examples, conclusion with a call to action.
Example of my style:
"""
[paste 2–3 lines of a text you wrote]
"""
You are a [language / framework] developer.
Create [feature] that [precise goal].
Constraints: [browsers/target], [no external dependency / use X], code commented in English.
Here is my current code so you follow my conventions:
"""
[excerpt]
"""
Deliver the complete code + a usage example, then list what you deliberately simplified.
Here is my text:
"""
[text]
"""
Correct it while keeping my voice (keep my wording when it is correct).
Type of correction: [spelling and style / complete restructuring].
Format: first the corrected version, then a table (before, after, reason) of the 5 main changes.
Here is my data: [file / table].
Question: [exactly what you want to know].
Format: [chart + 3 findings / table sorted by X / month-by-month trend].
Calculate precisely (no estimates). Point out missing or inconsistent data instead of ignoring it.
100% static guide: your quiz answers and checklist stay in your browser.