The first real request

Session: twenty minutes. You need: a writing task you have to do today anyway; if you have none, the lunch invitation from "Start here".

1. Pick one tool, just one

If your organisation provides one, use that: it's the one you'll be able to use with office documents. Otherwise open Claude (claude.ai), ChatGPT (chatgpt.com) or Gemini (gemini.google.com) and create a free account with your email. Copilot Chat, if your organisation uses Microsoft 365, opens with your work account from m365.cloud.microsoft/chat. For the first stages the choice of tool matters little: what you learn on one applies to the other. Switching tools every day, on the other hand, makes you start over every time. For all of level 1 the free plan is enough; what a subscription adds is in the fact sheet "The tools: free or paid", at the end of the guide.

Before you write: how the screen is laid out

The four interfaces look alike. At the bottom is the box where you type. Next to it, a "+" button or a paperclip to attach files and turn on tools such as web search. Near the box or at the top is the model selector: the same product lets you choose between different engines, faster and lighter ones or ones that can reason longer before answering. For today leave the default choice; in stage 2 you try changing it. Why there are models of different sizes is explained in chapter 3, in the sections "Big is not always better: efficient models" (page 72) and "New techniques for improving the capabilities of LLMs" (page 74), where you'll find reasoning models. Which models exist today, and who makes them, is on librovivoia.it/r/mercato.

2. Pick the task

It needs three features: it's real, it's for today, and if it comes out badly nothing happens. The reply to a supplier asking for an extension. The outline for Thursday's meeting. Three lines introducing a colleague to a client. If you teach, the notice to families for a school trip, with place, time and what to bring. If you study, the outline for presenting a topic you already know. If you need AI outside work, the message for organising something with other people. For now leave out anything that contains people's names, confidential figures or internal documents: you'll get there in stage 3.

3. Write the request as you would to a colleague who just started

A new colleague is capable, but knows nothing about you, the recipient or how you want the work done. The model is in the same position. Copy this template and fill in the brackets:

I work as [role] in [type of organisation]. I need to [task] for [recipient: who they are, what they already know, what they want from me]. The tone is [formal / friendly / plain]. I need [format: a ten-line email, a five-point list, an outline with timings]. Keep in mind that [the constraint an outsider can't guess].

A filled-in example:

I work as purchasing manager in a furniture company. I need to reply to a supplier who is asking for a two-week extension on delivering an order. We've known them for years and want to keep them, but the extension means a delay for one of our clients. The tone is friendly and firm. I need a ten-line email. Keep in mind that we can grant one week, not two.

4. Answer the answer

The first version is a draft. Read it and write what you would change, in your own words: "shorter", "the second sentence is too harsh", "the date is missing". The model rewrites taking into account everything you've said to each other. Do it at least twice: good work comes out of the second or third exchange.

Here's the difference on a real case: the same email, asked for first without context, then with the template from step 3, then corrected.

5. Actually use it

Reread the final text from start to finish, fix whatever you'd never say, and send it. With your name on it, you answer for it.

Two rules for today. Don't paste personal or confidential data. If facts, dates or numbers you didn't give it show up in the answer, check them before using them: stage 4 explains why.

What you just saw

You gave the model a context, a recipient, a format and a constraint, and you corrected the result by talking. These are the first techniques in chapter 5, "Getting what you need": open it at the section "Prompt engineering: how to talk to models" and you'll find what you just saw happen, with the reason.

In the book. Chapter 5, "Getting what you need", section 5.1 "Prompt engineering: how to talk to models", page 95; chapter 3, sections 3.2 and 3.3, pages 72 and 74, for models of different sizes. The book on Amazon.