You ask ChatGPT a question and get back something correct but bland, generic, only half usable. The problem is rarely the model — it's the instructions. A prompt is a brief. If you ask a competent person to «write me a LinkedIn post», you'll get an average LinkedIn post. With an AI, it works exactly the same way.
This page gives you a repeatable method: a 6-block structure to fill in order, a complete before/after example on a real case, the techniques that make the biggest difference, three ways to practice, and the most common mistakes. No technical skills required — writing a good prompt means clearly describing what you want, for whom, why, and in what form. Plan on ten minutes of reading, then an hour of practice to make it stick.
What Separates a Good Prompt From a Failed One
A good prompt contains everything a competent person would have needed to know before starting the work.
An AI model doesn't guess your intent. It generates the most likely continuation of your text, based on what you gave it. When your request is vague, the most likely continuation is also the most ordinary: an average text, written for nobody in particular. That's not a flaw in the tool, it's a mechanical consequence of missing information.
Here's what's almost always missing from a failed prompt:
- who the result is for (a client, your boss, a 10-year-old, just you);
- what it's going to be used for (sending an email, prepping a meeting, publishing an article);
- the expected tone (direct, warm, formal, technical);
- the length and format (5 bullets, a table, 300 words, an email subject line);
- the real information to use (numbers, names, context from your company);
- what the AI is not allowed to invent.
Compare these two instructions. «Summarize this text» produces a neutral summary of unpredictable length. «Summarize these meeting notes in 5 bullets max, one bullet per decision made, with the owner's name and the deadline in parentheses, without rehashing the discussion» produces something you can paste straight into an email.
One thing that often surprises beginners: length is not the criterion. A precise 40-word prompt beats a fuzzy 300-word one. What matters is the density of useful information.
The 6-Block Structure of an Effective Prompt
A solid prompt is always built in the same order: role, goal, context, constraints, format, examples.
| Block | The question it answers | Short example |
|---|---|---|
| Role | Who's speaking? | You're the customer service manager for an online store. |
| Goal | What result do I want? | Write a reply to a customer upset about a late delivery. |
| Context | What real info? | Order placed on March 12, 6 days late, cause: supplier shortage. |
| Constraints | What limits? | 120 words max, no promise of a refund, calm tone. |
| Format | In what form? | An email subject line, then the body of the message. |
| Examples | What should it look like? | Here are two replies we've already sent and approved. |
Role
Giving a role steers the vocabulary and the level of detail. «You're a middle school math teacher» and «you're a statistics researcher» won't produce the same explanation of the same formula. Stay realistic: a useful role describes a job and a level, not a superlative like «the world's best expert».
Goal
One action verb, one deliverable, one main task. «Analyze and write and translate and publish» in a single prompt gets you four sloppy jobs. Split them up.
Context
This is the block that makes the biggest difference, and the one people skip most. The AI doesn't know your company, your customers, or your constraints. Paste in your real data: email excerpts, numbers, notes, a page from your site. Without context, the AI fills the gaps by making things up.
Constraints
Length, tone, banned words, angle, what absolutely must not happen. Phrase your constraints positively when you can: «write sentences under 20 words» works better than «don't write long sentences».
Format
Describe the output like a mold: a 3-column table, a numbered list, JSON, a script with timestamps, an email with a subject line. If you're going to reuse the result in another tool, say so.
Examples
Two or three examples of what you're expecting beat a paragraph of explanation. It's the highest-return technique of them all, covered in detail in our guide to few-shot prompting.
Full Example: From a Shaky Prompt to One That Works
The same need, written twice: the difference comes down to six added lines.
The situation: you need to follow up with a client who never replied to a quote you sent two weeks ago.
The shaky prompt:
Write a follow-up email for a client who didn't reply to my quote.
Typical result: a polite, generic email full of phrases like «I'm just reaching out» and «please don't hesitate to let me know». Usable, but you're going to rewrite the whole thing.
The worked prompt:
You're a salesperson at a 6-person graphic design agency.
Goal: write a follow-up email to a prospect who hasn't replied to a quote.
Context: €4,200 quote sent on August 18 for the rebrand of a restaurant's visual identity. The contact is Julien, the owner. We spoke twice by phone, he was very interested, he was waiting on his business partner's approval. Our production slots for October are filling up.
Constraints: 100 words max, keep it professional with no overly casual tone, direct but not pushy, one single question at the end, no discount offer, don't write «I'm just reaching out».
Format: an email subject line under 50 characters, then the body of the message.
Give me 2 versions: one built around the scheduling constraint, one built around helping his partner decide.
This prompt gets you an email you can send after a thirty-second proofread. The blocks you added are the ones a colleague would have asked you about before writing.
Then you iterate instead of starting over. Round 2: «Version 1 is the right one. Shorten the subject line and replace the last sentence with a closed question.» Round 3: «Keep this email and write the next follow-up, to send 10 days later if there's still no reply.» Three short exchanges beat a perfect first-shot prompt every time.
The Techniques That Improve Results the Most
Four moves cover the vast majority of everyday use cases.
- Give examples. Paste in two or three past pieces you like, and say what you like about them (the rhythm, the structure, the level of detail). AI imitates better than it obeys adjectives.
- Ask for step-by-step reasoning. For a calculation, a comparison, a diagnosis, add «walk through your reasoning step by step before giving the final answer». The error rate drops noticeably on multi-step tasks. That's the chain-of-thought principle.
- Flip the questions around. End your prompt with «before you answer, ask me the 3 questions you need answered to do this job properly». You discover what you forgot to specify, which improves the next prompt.
- Break big tasks into pieces. A full article, a launch plan, a data analysis: ask for the outline first, approve it, then have it written section by section. Each approved step becomes the context for the next one.
The teams building these models publish their own recommendations, and they line up with this logic. Anthropic's official documentation on prompt engineering, for instance, emphasizes clear instructions, examples, and using tags to separate the parts of a prompt. For a broader survey of the techniques, our complete prompt engineering guide covers them one by one with examples.
Where to Practice Writing Good Prompts: 3 Methods
Progress comes from repetition on your real work, not from reading articles. Here are three ways to get started, from the most structured to the most independent.
Method 1: A Guided Program With the Tools Included (Skilzy)
Skilzy is a French e-learning platform built both for learning AI and for actually using it. The prompt fundamentals program takes you from the 6-block structure to reusable prompts for your own job, with graded exercises instead of theory.
What's different from a traditional course: the built-in AI Lab. You write your prompts and run them on real tools (text, image, video, music, voiceover) with credits included, without stacking up five subscriptions. The platform's other programs (more than 15 in total) cover image and video creation, automation with n8n, and AI-assisted coding.
Two state-recognized, funding-eligible certifications are available: RS7439, AI content marketing and RS6792, AI and sales. B2C access starts at €29.90 per month with no commitment, and there's a B2B offer for training organizations.
To try it without a credit card, the discovery demo gives you 7 days with 1 image, 1 video, 1 music track, and 10 messages. It's the only card-free entry point, and it's enough to see whether the format suits you.
Method 2: The Free Versions of ChatGPT, Claude, or Gemini
The three main assistants all offer a free tier with a daily message quota. It's the cheapest way to practice, and it works very well for text.
What you get: zero cost, instant availability, and the ability to compare answers from one model to another on the same prompt.
What you lose: no structured progression, no feedback on the quality of your prompts, quotas that cut you off in the middle of an exercise, and advanced features (image, video, large files) often reserved for paid tiers. If you're a genuine beginner, start with our guide on using ChatGPT as a beginner, then set yourself one work task per day for two weeks.
Method 3: Public Prompt Libraries
Free libraries exist from the AI companies themselves and on GitHub, with hundreds of ready-to-copy prompts. They're useful for getting moving fast on a task you don't know how to frame.
The limitation is real: a prompt copied without context gives a generic result, because it contains nothing about your company or your constraints. Use it as a skeleton, then fill in the context and constraints blocks with your own information. That's the difference between a prompt that works for its author and a prompt that works for you.
The Mistakes That Come Up Most Often
Most bad results come from five habits that are easy to fix.
- Stacking ten requests into one message. One task per prompt, then move on to the next.
- Only saying what you don't want. A list of prohibitions with no clear direction produces cautious, hollow text. Describe the target.
- Forgetting the audience. The same content for an expert and for a total beginner gives two very different texts, and the AI will default to the middle.
- Accepting the first answer. The first version is a draft to react to, not a deliverable.
- Trusting the numbers and sources it cites. An AI can produce a reference that doesn't exist. Check every figure, date, and link before publishing.
A sixth missing habit gets expensive over time: never saving your prompts. Keep a simple document with your 10 best prompts, sorted by task, and replace the variable data with brackets to fill in. These traps are covered in more depth in our article on the most common prompting mistakes.
Start Practicing Today
Take a task you handed to an AI this week and rewrite your request using the 6 blocks: role, goal, context, constraints, format, examples. Compare the two results side by side. It's the fastest exercise for understanding where the quality of an answer really comes from, and it takes no more than ten minutes.
