Measure, then wake up.

You described your plan and the assistant called it thoughtful, well reasoned and exciting. It was lovely to hear, until you wondered whether it says that to everyone.

It says it more often than a person would. A study by Cheng and colleagues, published in the journal Science in 2026, found that 11 leading AI models, asked for advice, affirmed the user’s actions about 50 percent more often than humans did. People in the study’s experiments also trusted the flattering models more and preferred them. The pull runs both ways: assistants lean toward agreeing, and we like being agreed with.

Measure is the M in DREAM, and it does two jobs. First, it tells the assistant how you will judge the answer: “A useful answer here is one that could talk me out of it.” Second, it asks the assistant to check its own work before handing it in: what it assumed, what it is least sure of, the case against, and where you need a source or a professional. A line like “where you don’t know, say so” gives it permission to be unsure. Anthropic recommends giving that permission outright, because it cuts down on made-up answers.

Then wake up

Measure improves the answer. It doesn’t make it true. So the last step happens on your side of the screen, and we call it WAKE.

• Where is this from? Ask for sources, and open at least one.
• Assumptions. Ask what it assumed about you and what was missing.
• Kick the tires. Find the one claim your decision rests on and check it somewhere else. For the bootcamp, that is whether people like you get hired afterward, and the place to check is two recent graduates, not the brochure.
• Escalate or edit. Health, money, law and safety go to a qualified person. Everything else, rewrite in your own words before you use it.

Side by side

The weak version ends in “right?”, which is a request for applause, and it will probably get some. The strong version shuts that door (“don’t cheer me on”) and says what a good answer looks like: one that could change your mind. It asks for both sides, then has the assistant audit itself and point to the one fact worth checking outside the chat. You still make the call. You just make it awake.

That is the whole anatomy: five parts for the prompt, four checks for the answer. The last lesson is about taking it to any app you like, and about the times the best prompt is no prompt at all.