The path is written around office work because that is where delegation is measured. The same moves apply outside it, and six uses deserve a line each, with the request to copy. Where one of the four tools behaves differently, I say so.
Studying a subject
The model does not know what you already know. Tell it with a concrete sentence, "I know percentages but not compound interest", and ask it to proceed one question at a time:
I am trying to understand the difference between weather and climate. Explain it to me as you would to someone who has not studied the subject, with an everyday example. Then ask me one question at a time and wait for my answer before moving on.
When a passage is not clear to you, say which one, rather than asking for "simpler". At the end close the chat, explain the concept in your own words and submit your explanation: "This is my explanation: [text]. What is right and what do I need to revise? If there is a mistake, lead me to it with a question before giving me the answer." A smooth text gives the feeling of having understood; the proof is answering a new question. If you teach, give age, objective, duration and prerequisites, and ask for six invented sentences with a task and a reasoned solution: check the examples and the solution yourself before taking them into class, and do not put real students' data into an exercise that works with invented names.
Speaking instead of writing
Two different things. Dictation turns your voice into text in the box: you can correct names, numbers and punctuation before sending. Voice conversation is a spoken exchange with the assistant, with the answers read aloud. In all four tools the microphone button sits next to the box, above all in the phone apps; if you cannot find it on the computer, use the dictation on your phone's keyboard. A well-dictated request: "I have to introduce myself in two minutes to a group of new colleagues. I will tell you what I do; help me put my ideas in order, and before writing the text ask me a question about what is missing." If you have spoken at length, at the end ask for a written summary and check it: speaking is easier than writing, not more precise.
Showing an image
All four read images: a photo of assembly instructions, a page of notes, the screenshot of an error. You attach it from the "+" like a document (stage 3). Frame it well and leave out passwords, codes and personal data that the question does not need. First ask for a transcription of the part you are interested in, "transcribe step 3; if a word or a symbol cannot be read, tell me", so you can see whether it is reading the right image, and only then for the explanation. An unclear photo gets retaken: a longer request does not recover what is missing from the image. Showing a screenshot does not give the tool control of your computer: reading what you show it and acting in your place are two different functions, and the second is stage 18.
Creating an image
ChatGPT and Copilot generate images. Claude draws charts and diagrams by writing the code that produces them (stage 8), but it does not generate an illustration: if you ask it for one, it prepares the description for you to use elsewhere. Describe subject, use and format: "A landscape illustration for the announcement of a reading workshop: adults sitting in a circle with books, in a bright library. Simple style, warm colours, free space at the top for the title. No lettering or logos." Then ask for one precise change at a time. Before publishing, check lettering and details, and do not use a generated image as evidence of a fact.
Translating
Say where you will use the text, because an email to a client and a public notice call for different choices, and ask it to flag ambiguities before resolving them: "Translate this email into Italian for a colleague I am on friendly terms with. Dates, amounts and commitments stay exactly as they are. If a sentence in the original is ambiguous, ask me before choosing a meaning." Check names, numbers, negations and deadlines. If you do not know the language, ask for an explanation in your own language of the most important choices, knowing that it is an aid to checking and not a verification: for a text on which a commitment depends, the rereading is done by a person who knows the language.
Bringing out alternatives and comparing them
"I have to organise a reading meeting for twelve people, in a small room, one hour, no projector. Propose three different activities; for each, preparation, materials and one possible difficulty. Then ask me which one to develop." If the three look alike, change a criterion: more participation, less preparation. To compare two proposals, you supply the criteria and the data, and ask for "to be verified" where a figure is missing, with no winner until you have said which criteria weigh most. The comparison should show the reason for the choice, not produce a ranking.
In the book. Chapter 5, section 5.1 "Prompt engineering: how to talk to models", page 95, for the role, the examples and the constraints that apply here too; chapter 7, section 7.1 on hallucinations, page 118, for checking translations and images read. The book on Amazon.