Treat the first answer as a draft

The doctors’ trial has one more lesson. The researchers didn’t write their prompt once; they refined it over several rounds.¹ A first answer is a first draft, and the request that produced it can be a draft too. Three habits help you improve both.

• Hand it your sources. When the facts are in front of the model, it doesn’t have to recall them. In a 2020 study, a model that looked up relevant documents and read them along with the question gave more specific and factual responses than one without that step.² Paste the policy, attach the schedule, quote the email you are answering.
• Ask it to ask you. Begin with “Before you start, ask me anything you need to know.” Its questions show you what it would otherwise have guessed.
• Say what to change. A follow-up such as “shorter, and lead with the deadline” takes a few seconds and keeps what already works.

Once a request works, keep it. The trial’s authors suggested that organizations could build well-designed prompts into the tools doctors already use, so that nobody has to invent a good one from scratch.¹ You can do a small version of the same thing. Save the requests that worked for tasks you repeat, such as the monthly summary or the weekly update, and next time change only the parts that differ. A saved request is also easy to share with a colleague, which spreads the asking along with the answer.

An illustration, not a real case. A middle-school teacher is about to type “a quiz on fractions.” She stops and runs through the six parts instead. Who it’s for: sixth graders who have just learned to add fractions with different denominators. An example: one question copied from her own worksheet. Limits: eight questions, no word problems. The shape: questions first, answer key at the end. When the draft comes back, she reads it as she would a student teacher’s, then replies with what to change: “Make the last two easier, and add one with mixed numbers.” Two short rounds, and the quiz is hers.

This book was made the same way. Its chapters were drafted with AI help from instructions that named the reader, the research standard and the examples to use. Then every claim was checked against its source, and the drafts were revised.

Your turn

Try it. Pick a real task from your week: an email, a summary, a plan. With any AI tool you already use, send your usual one-line request first and keep the answer. Then start a new conversation and send the six-part version, with the key instruction at the start. Compare the two. Then test one of this chapter’s claims: begin a third conversation with “Before you start, ask me anything you need to know,” and count how many of its questions your first request had left it to guess. On paper instead: write out the six parts for the task, then circle each part your usual request leaves out. Each circle is a guess you have been asking the model to make.

Think about it. Think of the best briefing you ever got at work, from a manager, a mentor or a colleague handing over a project. Which of the six parts did it include? Which do you tend to leave out when you hand work to someone else?

So what makes one request work better than another is what makes a good briefing work: it says the job, who it’s for, the background, an example, the limits and the shape, and it puts the key instruction where it won’t get lost. Then it treats the first answer as a draft. A better request gets you a better draft, but not a checked one. The doctors’ trial ended with a warning against trusting the model without oversight, and that raises the next chapter’s question: how do you check what an AI tool tells you?