Locked learning resources

Join us and get access to thousands of tutorials and a community of expert Pythonistas.

Unlock This Lesson

Locked learning resources

This lesson is for members only. Join us and get access to thousands of tutorials and a community of expert Pythonistas.

Unlock This Lesson

Asking Well

Sarah Jane Ross
AI for Beginners: Use Agents With Confidence Sarah Jane Ross 03:32

Coming back after a break? Take a five-minute warm-up on Sections 1 and 2 before you start.

The words you send an AI assistant are called a prompt, and the way you ask shapes what comes back. Asking well comes down to being clear about what you want and handing over what the model can’t know on its own.

What a Vague Request Gets You

The model knows only what’s in the context. When a request leaves something out, the model fills the gap with something typical. Ask for “an announcement about our new expense tool,” and you get a polished, generic announcement, with blanks for the date and the tool’s name and a guess about the training.

That reply is the most likely answer to a request that could mean many things.

What Makes a Request Better

Locked learning resources

Join us and get access to thousands of tutorials and a community of expert Pythonistas.

Unlock This Lesson

Already a member? Sign-In

Locked learning resources

The full lesson is for members only. Join us and get access to thousands of tutorials and a community of expert Pythonistas.

Unlock This Lesson

Already a member? Sign-In

00:00 Asking Well.

00:02 The words you send an AI assistant are called a prompt, and the way you ask shapes what comes back. Asking well comes down to two things. Be clear about what you want.

00:12 And hand over what the model can’t know on its own. The model knows only what’s in the context. When a request leaves something out, the model fills the gap with something typical.

00:24 Ask for an announcement about your new expense tool, and you get a polished, generic draft. It has blanks for the date and the tool’s name. And it makes a guess about the training.

00:35 That reply is the most likely answer to a request that could mean many things. Now add what you know. Say it’s for the company team chat, in under sixty words.

00:46 Then give it the date, the tool’s name, and the training plan. The draft gets shorter, and the blanks fill in.

00:54 Most of what improves a result comes down to four things. First, say what done looks like. Who is it for, and what is it for? How long should it be, and in what format?

01:05 Second, give it what it can’t see. That means your situation, the facts, the names, and any files it needs. The model can’t look these up in your head or your inbox.

01:16 Third, show an example. If a message worked last time, paste it in. An example shows the tone and length you want at a glance.

01:25 Fourth, follow up. Treat the first reply as a draft. Read it, then say what’s off. Maybe it should be shorter, or less formal. Maybe it left out the deadline.

01:36 Each follow-up builds on the conversation so far. When you’re not sure what you want yet, you can turn it around. Ask the assistant to ask you questions first.

01:46 Its questions show you what’s missing from your request. An agent that works on its own over many steps is different. You can’t follow up after each step.

01:54 So the pieces that matter have to be in the request from the start. The goal, which is what you want at the end. The limits, like a budget or a date range.

02:04 When to check with you, like before it buys, sends, or deletes anything. And what done looks like, meaning what it should hand back, and in what form.

02:13 Say you ask an agent to find you a hotel for the conference. It would have to guess almost everything. Which conference, which dates, and what you’re willing to pay.

02:22 Whether it should book, or only look. A better goal asks for three hotels within a fifteen-minute walk of the venue, for two named nights, within your company’s travel limit, with free cancellation.

02:33 It says not to book anything. And it asks for a short table with the price, the walking time, and the cancellation deadline.

02:41 Many assistants also accept a photo or a screenshot, and some let you talk instead of type. The same rules apply. A screenshot of an error message, with nothing but a short question, leaves a lot to guess.

02:54 Say which app it came from, what you were doing, and what you’ve already tried. You’ll get a much better answer. As models get better, you’ll need to steer them less.

03:04 Tricks that once helped, like special phrases, matter less and less. But two things stay the same, however good the models get. Only you know what you want.

03:13 And the model can see only what you give it.

03:17 Below this video, there’s a practice app. Start with the one-line request, and read what comes back. Then improve it one piece at a time, and watch the draft change.

03:26 You can also start a new chat, and let the assistant ask the questions.

Become a Member to join the conversation.