Most people who say "AI doesn't understand me" are describing a gap in their own message. Chatbots do not read minds and they do not look things up about you. They work with what you type. A few extra lines of context, added in the right order, change the result more than switching to a more expensive subscription.
What a prompt actually is
A prompt is not a search query. A search engine guesses from three words. A chat model reads every sentence you give it and builds on all of them. It still knows nothing about you: not your job, not who will read the output, not what you rejected five minutes ago.
That explains most of the disappointing answers. The model is not failing to understand English. It is filling the gaps you left open, and it fills them with the most average, safest version of whatever you asked for.
Five parts of a prompt that works
Your role and the reader
Start with who you are and who the text is for. "I run a small accounting office in Plzeň" tells the model more than "help me with email". So does "the reader is a client who has never heard of VAT reverse charge". Two sentences of this do more than a page of instructions written in technical shorthand.
The context around the task
Say what led to the request. A teacher writing to parents about a cancelled trip needs a different tone than a manager announcing a restructure. If you have tried something already and it did not work, say so. "Last version was too formal" is useful information.
The shape of the answer
Do you want three bullet points, a table, a 150-word paragraph, a numbered procedure? If you say nothing, you get whatever length feels natural to the model, which is often longer than you need. Say "maximum 120 words" and you will get about that.
One example
Paste an example of the style you want. One short sample teaches more than a paragraph of description. This works for product descriptions, replies to complaints, even the way you write a shopping list.
The limits
Tell it what not to do: no jargon, no emoji, British spelling, do not invent figures, ask me a question if something is missing. The last one is worth remembering. A model that asks you a question is usually about to give a better answer.
The same request, written two ways
| What you usually type | What to type instead |
|---|---|
| Write an email about the price increase. | Write a 120-word email to our 40 small business clients in Czechia explaining a 6 percent price increase from January. Tone: apologetic but firm. No marketing language. End with the name of our support line. |
| Summarise this contract. | Summarise this contract for someone with no legal training. List the payment terms, the notice period and the cancellation conditions as three bullets. Point out any sentence that is unclear rather than guessing what it means. |
| Ideas for a poster. | I design a poster for a small bakery in Olomouc. Give me eight visual concepts, each two sentences, using only materials a screen printer can handle. Avoid anything with photo montage. |
None of these examples is clever. They are just specific. That is the whole method.
Correcting the answer
Do not start over when the answer is wrong. Reply with what is wrong and what should change. "The second paragraph is too aggressive, keep the rest" gives the model something to work with. Restarting a fresh chat throws away the context you already built.
Two corrections are usually enough. If you are on your fourth, the problem is probably the task itself. Split it: ask for the structure first, then fill in the parts one by one.
Language and the European detail
If you work in Czech, Slovak, German or Polish, write your prompt in that language when the output has to be in it. All the major chatbots operate in Czech and will reply in Czech, though the quality of a long legal or medical text in Czech is usually below what the same model produces in English.
A practical habit: draft in English, then ask for a translation and a review by someone who speaks the language. Machine translation into Czech still leaves gender endings and word order that a native reader spots at once. The French company Mistral runs Le Chat, which is free and available across the EU, and it is a reasonable choice if you want a European provider. Coverage of smaller EU languages is uneven across all providers, so test with your own text before you rely on it.
What not to paste in
Under the GDPR, a chat message that contains a pupil's name, a patient's diagnosis, a client's contract or a staff member's review is processing personal data. You are the one responsible for it, not the chatbot.
Consumer tiers are a weak spot. Free and individual paid plans usually reserve the right to use conversations for training unless you turn that off in the settings, and the settings differ by provider. Business and team plans typically offer different terms and a data processing agreement. If you work at a school, a clinic or a company, check with whoever handles data protection before you paste anything real.
The EU rules are already in place. The AI Act's obligations for general-purpose AI models, the kind behind chatbots, have applied since 2 August 2025. GDPR has applied all along. In practice that means two things for an ordinary user: keep personal data out of consumer chats, and do not pass off machine-written text as the work of a named author without checking it.
Free tiers and paid plans in Europe
All the major assistants are available in the EU. ChatGPT, Claude, Gemini, Copilot and Le Chat all have free tiers that are good enough to learn on. Paid plans sit around $20 a month for the individual tier, which is roughly €18 to €19 before VAT at current exchange rates. European customers see VAT added at checkout, so the final figure is closer to €22 in Czechia or Germany and higher in Hungary.
| Service | Free tier | Individual paid tier |
|---|---|---|
| ChatGPT | Yes, with usage limits | About $20/month, VAT added in the EU |
| Claude | Yes, with usage limits | About $20/month, VAT added in the EU |
| Gemini | Yes | Around $20/month through Google's plan |
| Mistral Le Chat | Yes | Paid tier available, priced in euros |
Prices and plan names change; check the provider's page before you subscribe. For learning to write a prompt, the free tier is enough. Paying buys longer conversations, newer models and fewer limits, not better judgement on your side.
A short routine that saves time
Before you press send, check four things. Does the model know who you are and who reads the result? Have you said how long the answer should be? Is there an example or a format you can point to? Have you removed the names and the account numbers?
Ten extra seconds of typing, most days. The gain is not a perfect answer. It is an answer you can fix in one round instead of five. Our AI courses cover this in more depth if you want the longer version.
A prompt can organise what a model already knows. It cannot supply a fact the model never learned, and it cannot turn a wrong number into a right one. Check anything that ends up on a patient's chart, in a signed contract, or in front of a class.
Do I need to pay for a subscription to write good prompts?
No. The free tiers of ChatGPT, Claude, Gemini and Le Chat all handle the basic work described here. Paid plans add longer context, newer models and fewer limits. A clear prompt on a free plan usually beats a vague prompt on a paid one.
Can I reuse the same prompt in different chatbots?
Often yes, and it is a fair way to compare them. Results will differ because each model is trained differently and each has its own default length and tone. Keep the prompt the same, change one variable at a time, and judge with your own material rather than a demo.
What is a prompt library and is it useful?
It is a saved collection of prompts for tasks you repeat, such as meeting notes, product descriptions or replies to complaints. It saves typing, but a saved prompt still needs the context of the day. Treat it as a starting template, not a fixed recipe.
