Text assistants: pick one, use it well

There are three general-purpose assistants that matter for everyday work, and arguing about which is best is the most popular way of not using any of them. This lesson settles it in a paragraph, then spends the rest of the time on the thing that actually changes your results: how you ask.

Pick one. Here's how.

ChatGPT, Claude and Gemini are all good enough for ordinary work: writing, summarising, rewriting, explaining, planning, first drafts of almost anything. They differ at the edges, and the edges move every few months, so any ranking you read is already stale.

The decision procedure:

1. If your company allows only one, use that one. Microsoft Copilot counts if that's what you've got; it's built on the same kind of model.
2. If it allows several, use the one you've already opened. You'll learn more from a month with one than a week with each.
3. If you've opened none, pick Claude or ChatGPT, flip a coin if you must, and don't revisit the choice for a month.

The skill you're about to build transfers. If you switch later, you lose nothing.

Should you pay?

Not yet. The free tiers do the work in this course. Pay when, and only when, the free version blocks something you're already doing regularly: you keep hitting the message limit on real tasks, you need to attach big files, or you need the stronger model for something where the difference is visible to you. "I might need it" is not a reason. One subscription is enough; two is a hobby.

The shape of a request

Here is where almost all the difference between "it's useless" and "it's great" lives. Remember from lesson 1: the model can only see what's in the window, and it produces whatever shape of text seems most plausible. So a one-line request gets a plausible generic answer. A request that carries the context gets an answer for you.

Walk through them:

Who you are. One clause. "I'm an operations manager", "I'm writing as the HR lead", "I run a two-person design studio". It sets the register and the assumptions. Without it, the model writes for an imaginary average person.

The task, and why. Not just "write an email about the delay" but what the email is for. Knowing the purpose lets it make the hundred small trade-offs you'd make yourself.

Who reads it, and how. On a phone, in a hurry, between meetings. Or carefully, with a lawyer. This single line decides length, structure and tone better than any instruction about "professional but friendly".

The material. Paste it. Don't describe it. "Here's the internal update:" followed by the update beats "based on what we discussed" every time, because the model can't see what you discussed. Attach files when they're long. The facts in the reply should come from your material; anything it adds on its own is the part you check.

The shape of the output. This is the one people skip and then complain about. Say what it should look like: one email under 120 words; a table with three columns; five bullet points, each one line; a two-paragraph summary followed by three open questions. Describing the shape, not just the length, is the fix for "it writes an essay when I wanted two lines." The guide Why AI ignores your instructions on this site has eight habits for this; the short version is that you describe what the result looks like, including what to avoid.

Follow up. Don't start over.

The first reply is a draft. The second request is where it gets good:

• "Shorter. Cut the second paragraph entirely."
• "Keep the structure, make the tone plainer. No 'I hope this finds you well'."
• "Now give me a version for the client who's already angry."
• "Which of those three points is weakest, and why?"

Each follow-up is cheap, because the material is already in the window. Three quick follow-ups usually beat one perfect prompt. Resist the urge to rewrite your original request from scratch; just say what to change.

One chat per task

Start a new chat for each new task. The window stays small, nothing leaks across tasks, the assistant isn't weighing Monday's poem while writing Thursday's report, and you can find things again. Most products let you name or pin chats; "Client delay email" is a better name than "New chat 14".

The one exception is a genuine continuation: you're still on the same document, the same analysis. Then stay. But "same week" is not "same task".

What it's reliably good at today

For someone whose job isn't AI, these are the safe bets. Start here, not with the exotic stuff.

• First drafts of emails, documents, posts, proposals. You edit; it starts.
• Rewriting: shorter, plainer, more formal, for a different reader.
• Summarising long things you pasted, and listing the questions they leave open.
• Explaining something you half-understand, at the level you ask for.
• Structuring: turning a messy brain-dump into an outline, an agenda, a plan.
• Checking your own writing for gaps, unclear bits, and what a sceptical reader would ask.

And where to stay alert: anything with specific figures, names, dates, law, medicine, or money, it's a draft that needs checking against a real source. More on that next lesson.