Text and chat
A chatbot can write, explain, translate and come up with ideas. See it as a rough draft. It writes the first version, you make it final. Learn four steps to get a draft worth working on.
First, what do you think?
You type "write an email to my landlord" into a chatbot. Will the email be any good? Have a guess.
What can a chatbot do with text?
The whole story in three sentences. A chatbot predicts text, so text is what it's best at. It can write, explain, translate and come up with ideas. What it writes is a rough draft, not the final version.
Think of a rough draft. It's quick, and it gets you started. But you wouldn't hand it in like that. You read it, fix it, and make it yours.
Google lists what language models do. They write sentences, paragraphs or poems. They answer questions, translate, and carry on a conversation.
- Writing. A letter, an invitation, a story, a speech.
- Explaining. "Explain photosynthesis like I'm ten." Then ask follow-up questions.
- Translating. A text from Dutch to English, or the other way round.
- Ideas. Ten names for a club, five ideas for a birthday party.
What all four have in common: you get a draft. Google also writes that the predictions often contain mistakes. So you're the editor.
What the predictor does
The predictor continues what's on the whiteboard. That's the rule.
A short question gives it little to go on. "Write an email to my landlord" could be a hundred different emails. Friendly or angry? About a leak or about the rent? The predictor fills the gaps with what's most common. That's usually bland.
A question with details gives it more to work with. Who you are, what happened, what you want, and how long. Then the most likely next words fit your situation.
Google's own advice for prompts says the same. Give clear and specific instructions. Add the information the model needs. Say what form you want. And expect to try a few times.
The more you put on the whiteboard, the better the draft fits. And it's still a draft.
Try it yourself
Three tasks, two ways to ask. Which question gives the predictor more to work with?
Everything runs in your browser. Nothing you type is sent anywhere.
How do you get a good draft?
Four steps. Try them today on something small, like a message to a friend.
Say what you want
Name the task, who it's for, and how long. Not only "an email to my landlord". Better: a short, polite email about the tap that has leaked for a week.
Give context and an example
Add what the predictor can't know. Your situation, the facts, earlier messages. If you have an example of the style you want, paste it. Google recommends examples in prompts.
Say what form you want
A list, a table, three sentences, a letter. Under 100 words. In simple language. Google calls this the response format.
Improve in rounds
The first answer is rarely perfect. Say what to change. Shorter, friendlier, without the second paragraph. Then read the final version yourself, and change what isn't you.
In practice
Three people get a draft. The steps make the difference.
Case 1The talk about dolphins
Sanne (11) has to give a talk about dolphins. She asks a chatbot: "Write a talk about dolphins."
Read out what comes back.
It isn't her talk. Schools agree that AI helps, but doesn't replace your own work. And the facts may contain mistakes.
Ask: "Explain how dolphins sleep, for a ten-year-old." Then ask follow-up questions. Check the facts in a book, and write the talk herself.
The chatbot explains. Sanne learns, and the talk is hers.
Case 2The letter to the council
Mehmet (42) wants to object to a parking fine. He finds it hard to write formal letters.
Not send anything, or send an angry email.
Without a clear letter, the objection may go nowhere.
Give the chatbot the facts: date, place, why the fine is wrong. Ask for a short, polite letter. Then check every fact and fill in his own details.
The draft gives him structure. The facts and the decision stay his.
Case 3The patient summary
Esther (35) works at a doctor's practice. She wants to paste a patient letter into a chatbot to summarise it.
Paste it in. It saves time.
The Dutch data protection authority got reports of exactly this. Patient data was typed into a chatbot. That counts as a data breach.
Only use tools her employer has approved for patient data. Or leave out all personal details first.
The patient's data stays where it belongs.
When not to
- Don't use it as your source of factsPredictions often contain mistakes.Use it to write and explain, and check facts in a source.
- Don't hand in its work as yoursAI helps, but doesn't replace your work.Use it for ideas and feedback, write it yourself, and be open about using it.
- Don't paste private informationNames, addresses, medical or customer data.Leave it out, or use a tool that's approved for it.
By situation
Use only the tools your school allows, and say when you used AI. Ask it to explain, to quiz you, or to give feedback on your text. Writing it yourself is how you learn.
Drafts for emails, reports and texts save time. Check every fact and figure. Never paste customer or patient data into a tool your employer hasn't approved.
A letter to the council, a complaint, an invitation, a recipe. Give the details and ask for the form you want. Read it before you send it.
Myths and mistakes
| Claim | True? | Why |
|---|---|---|
| A short question is enough | Not usually | The predictor fills gaps with what's common. Details make it fit. |
| The first answer is the best one | No | Improving in rounds is normal. Google says so too. |
| If it writes it, it's correct | No | Predictions often contain mistakes. You're the editor. |
| Using AI for homework is always cheating | No | Many schools allow it for help, as long as you're open and do your own work. |
Related topics
Read first
- From model to assistantthe layers behind the chatbot.
Next
- Imagesmaking and editing pictures with AI.
Related
- A good promptmore about asking the right question.
Sources
- explainerGoogle for Developers. LLMs: What's a large language model?.A language model predicts the next token, like a very large autocomplete. A token can be a word, part of a word or one character. Training needs a huge amount of text. Predictions often contain mistakes.
- documentationGoogle AI for Developers. Prompt design strategies.The model picks the next token with some chance, set by the 'temperature'. Low: more the same each time. High: more varied answers. Prompt advice: give clear and specific instructions, add context, include a few examples, say what format you want. Not the result you expected? Rephrase it; a good prompt often takes a few tries.
- organisationKennisnet. Schoolafspraken over het gebruik van generatieve AI.Principles for schools: people come first, use only tools the school approved, be open about using AI, AI helps but never replaces your own work, protect privacy. Under 13 only with an adult.
- governmentAutoriteit Persoonsgegevens. Caution: use of AI chatbot may lead to data breaches.The Dutch data protection authority got reports of data breaches: a GP practice employee typed patient data into a chatbot, a telecom employee a file of customer addresses. Most companies behind chatbots store everything you type. (Checked through a search engine.)
Test
You get 7 questions about the whole page. Get 5 right, and you've mastered the topic.
About the tools in this course. Every tool is described in the same way, in alphabetical order, without a recommendation. Names, prices and features change often: the maker's own site is always right. This site was built with the help of Claude, one of the tools described here.