Getting useful work out of an AI chat

Better results start before you open the chat. Five minutes spent writing down what you need gets you further than a conversation spent working it out. The rest comes down to a few habits for asking well, checking answers and protecting what you share.

In one large study, half of all AI chat conversations ended after two requests or fewer,¹ so in many chats the first request is one of only two, or the only one.

Typed straight into the chat, a request carries only what came to mind in that moment. The AI fills in the rest on its own.²

Instead of starting to chat, it’s more effective to prepare the request outside the chat. You have time to think it through, and the AI has less to guess. Its first answer will come closer to what you need, and you’ll spend less of the chat correcting it.

Pro tip: draft the request outside the chat

Write it in a notes app, a document or an email draft. This is your note. Paste it into the chat once it’s ready. Preparing it first makes you decide what you want before the AI decides for you. You can also adjust the note and reuse it for similar tasks.

Six things the note needs

Write down six things. Unless you tell it, the AI has to guess them.^3,4

• What you want done. Write the task as a clear instruction, such as “rewrite this email” or “list the risks in this report”. The AI does what the words say, so “suggest changes” gets you suggestions and “rewrite” gets you a new version.
• Who it is for, and why. Say who will read the result and what it should achieve. A summary for your board and one for a new colleague read very differently, and the AI cannot tell which one you need.
• What it should look like. Give the length, the tone and the format, for example “five bullet points” or “a short, friendly email”. Otherwise you get the AI’s default answer.
• Your material. Add the draft, the document or the facts you have gathered. Include two or three examples you like as well, because an example shows the tone faster than a description.
• Where the facts come from. Say which facts the AI may use, such as your document or a source you name. Ask it to flag anything it can’t confirm.
• What must stay as it is. List the dates, prices, names or approved wording the AI must leave unchanged.

Give it something to work on

AI does well when it edits text you give it. In data from one AI company, improving an email, rewriting a short message and correcting spelling each succeeded more than 94 times out of 100, against about 70 for all tasks. Writing a whole academic paper succeeded about 6 times out of 100.⁵ A draft or a document turns a request to invent into a request to improve. With a long document, put it first and your question last.^4,6

Check anything current online

An AI model’s knowledge stops at a fixed date, and the model is usually released months later.

A model released in September 2026 knows nothing after April 2026.^7,8

The gap grows every month a model stays in use. In the same company’s data, questions about recent company news or league tables succeeded fewer than 5 times out of 100.⁵

For anything current, give the AI today’s date and tell it to check online.⁶

Where the chat can search the web, that instruction makes it more likely to look the facts up.⁴ Then open the sources it cites.

Plain words work better than pressure

Adding “critical” or “always calculate correctly” does not make the maths right.⁹ Some newer models overreact to that kind of wording.⁴ For numbers, the AI needs a way to calculate, and for recent facts it needs a source.⁶ Ask for those in plain words.

Build on the answer, or restart from the note

Once the AI has answered, work with what you’ve got.

Ask it to change the parts that miss the mark and keep the rest.

Tasks that people work through with the AI end well more often than tasks they hand over whole.¹⁰ Experienced users of AI chats work like this more than newcomers do, and more of their conversations succeed.¹¹

An answer you have to throw away costs more time than checking a good one.

You spend the time reading it and still have to start again.¹⁰ Answers vary between attempts, so if asking for changes doesn’t help, try the same request once more before you rewrite it.⁹ When a conversation has gone wrong from the start, add what was missing to your note and open a new chat.

Know what you hand over

Everything you paste into an online AI chat goes to the company that runs it. At one provider, conversations on personal accounts, paid plans included, are used for training by default until you switch this off.¹² There, rating an answer can still send the whole conversation for training.¹²

Its business accounts are excluded from training by default,¹² so if your employer has an account like that, keep client work there.

One company keeps rated conversations for five years.¹³ Another keeps chats read by its human reviewers for up to three years, even after you delete them.¹⁴

So, here’s how it looks in real life: the same request, before and after.

Before: “Write an email to a client about a delay.”

After, in the note:

• Task: an email telling the client that delivery is moving from 12 to 19 November.
• Reader: their purchasing manager, who has planned around the 12th.
• Goal: keep the order, with part of the delivery offered on the 12th.
• Tone and length: polite, direct, under 120 words, one line of apology.
• Attached: our last two emails to this client, to match the style.
• Facts: use only the dates and terms in our last email.
• Fixed: the email gives no date later than the 19th.

Every line adds something the first version left the AI to guess.

What to remember before your next AI chat

1. Write the note before you open the chat.
2. Attach your material, and give the AI a source for anything current.
3. Ask in plain words, then build on the answer or start again from a better note.
4. Before you paste client work, check whether your account lets the provider train on it.

A task the AI got wrong cost about 20 minutes,¹⁰ while the note takes five, so it’s worth trying