First dialogue: how to talk to ChatGPT

The account exists, the app is installed, the settings are yours. Time to talk — and this is where almost everyone makes the same mistake. A person opens the chat and types the way they've typed into search engines for twenty years: "holiday turkey september cheap". That was enough for a search engine, because it matched pages to words. An AI is built differently: it doesn't search, it answers — and it answers exactly as well as you explained what you want.

Feel the difference on one task

One vivid experiment works better than any list of rules. Take a task everyone faces sooner or later: a present for your mother's birthday.

The search-engine request — "gift mum 60" — brings back the internet's average: perfume, a scarf, a voucher. Google would have given you that; no AI required.

Now the same question told like a human being: "My mother is turning 60. She's just retired, misses her school job, adores her garden and her grandchildren, and says she has everything. Budget €100–150. Suggest 10 ideas — warm ones, nothing generic." The answer changes beyond recognition: a floristry workshop, a photo book with her students, rare rose varieties. Ideas not for "a mother" but for your mother.

Same tool. The whole difference is how much you told it.

Three rules everything else rests on

That experiment boils down to three rules. Everything later in the course is built on top of them.

1. Write as if to a person. Full sentences, with details: "Help me plan an inexpensive holiday in Turkey in September for two, budget €1,500." The model picks up the budget, the party and the season — it doesn't need keywords, it needs the picture.
2. Remember: it remembers. Within one chat the model sees the whole conversation, so you can write "now shorter", "replace the second point" — it knows what you mean. For a new topic, open a new chat so the old conversation doesn't get underfoot: more on that in the lesson about tidy chats.
3. Ask for another take. The first answer is a draft, not a verdict. "Too formal, make it simpler", "add examples", "cut it in half" — every edit moves the result closer to what you actually need.

What it looks like live

Here's a real conversation, three moves. Notice that the person never repeats the setup — the model keeps it in mind by itself.

You: Write a listing to sell a sofa, 2 years old, great condition, €200.
AI: [produces three paragraphs of "luxurious" and "stunning"]
You: Too salesy. Calmer, shorter, just the facts. And add: buyer collects.
AI: [produces 4 honest sentences]
You: Good. Now the same thing for the neighbourhood chat — fully informal.

Three moves — three versions of the listing for three different places. Typing those replies took under a minute.

What a strong request is made of

Take good requests apart and you find the same three parts inside: the situation, the constraints and the result you want. Watch what happens when those parts appear.

• "write to the client" → "Write a short email to the client: we delayed the delivery by 3 days because of customs; apologise without grovelling and offer 5% off the next order"
• "explain inflation" → "Explain inflation to a 12-year-old through pocket money and the price of ice cream"
• "suggest a name" → "10 names for a coffee shop by the station: short, easy to catch by ear, without the words 'coffee' and 'bean'"

That trio is worth memorising: it comes up again and again, and half the craft rests on it.

If the answer misses

A poor answer is not a reason to decide "the AI is stupid" or "I can't do this". The cause is almost always one of three, and each is fixed with a single sentence.

• Too little context — add the situation and the goal: who it's for, what matters, what's off limits.
• A vague ask — replace "make it better" with a concrete verb: shorten, soften, find the mistakes.
• Wrong shape — say what the result should look like: a list, a table, three paragraphs.

Three fears you can say goodbye to

• "I'll ask something stupid" There's nobody to laugh. It's a program; it keeps no list of "stupid users" and won't tell your colleagues.
• "I'll break something" There's nothing to break. The worst that can happen is a bad answer. Ask again.
• "It won't understand my language" It will, fluently. Write however is comfortable, typos included — modern models cope with far worse.

Why one question gets different answers

One last detail that trips people up. Ask the same question twice and the answers will differ. That's not a fault: the model composes each answer afresh, with a share of randomness. The setting has a name — "temperature" — and its whole purpose is to keep answers from being identical.

It's easy to turn from a quirk into a tool: didn't like the answer — ask for "another take", and you'll get a genuinely different one, not the same again. Three options to choose from beat one almost every time.

Do this now

Take something real from the next few days — an email, a plan, an explanation — and write the request using the trio from this lesson: situation, constraints, result. When the answer comes, ask in one sentence for a redo. That's the basic loop all AI work is made of: draft — edit — done.