Start with a specific outcome, show what good work looks like, and test the result outside the chat. For factual answers, verify the source and what it actually supports. A confident answer—or a second AI agreeing with it—is not verification.
Companion guide · Updated · AI authorship & limitations
A fluent answer can still miss the point. To get better work from AI, give it a destination it can test, then inspect the result yourself.
1. Start with the outcome
Describe the person you want to help and what should be different when the work succeeds. “Write something” leaves the goal invisible. “Help a parent compare two evidence-based ways to protect sleep without losing important connections” gives the assistant a useful direction.
2. Show the standard
Give a good example, a rejected example, or a short explanation of what you would consider a failure. Hidden preferences become visible when you say why an output did not work. Ask the system to list what it still needs to know rather than filling gaps with confidence.
3. Test the result outside the chat
Open the page. Try the game. Check the cited paper. Ask another person to use the product without coaching. An AI-generated answer is a proposal until the intended person can use it and the important claims survive inspection.
4. Let the system challenge your idea
Invite a counterexample. Ask which assumption is doing the most work. If the answer concerns health, money, law, or safety, use appropriate qualified sources and expert review. An AI critic can expose a weakness, but it can invent one too—check the criticism against evidence.
Use the guide to writing with AI while keeping your voice, including a free brief builder and a feedback-only prompt.
“I want to help [person] achieve [outcome]. Here is the context and a good example. Tell me what is missing, propose two approaches, state what would disprove each, and give me the smallest real-world test.”
That approach is one practical companion to the book’s larger question: is the answer merely attractive, or did it help us get somewhere worth going?
How do you check an AI answer?
Use this five-step checklist on one important claim before relying on a long answer. It is a practical exercise from this companion site, not a certified evaluation procedure.
- Isolate the claim. Write one sentence that could be true or false. Separate a reported fact from advice, an analogy, or a prediction.
- Open the source yourself. Check the title, author, date and publisher. A plausible citation that you cannot locate remains unverified.
- Find the supporting passage. Does it establish this exact claim, or merely discuss the same topic? Record the section or page so someone else can check.
- Check the scope. Who or what was studied, what was compared, and for how long? A result about one task or group does not automatically apply everywhere.
- Test the useful part. Recalculate a number, run the proposed code on a small safe example, or ask a person to try the instructions. Keep unresolved uncertainty visible.
For the wider risk picture, consult NIST’s Generative AI Profile. The checklist above is our suggested workflow; it is not a NIST endorsement or certification.
A worked example: “the turtles follow the moon”
That sentence is an attractive summary, but it leaves out an important detail. Ask whether hatchlings need the moon to be visible. Then inspect the wildlife-agency explanation of the brighter seaward horizon and darker landward outline. The sea-turtle lighting guide shows why a compelling game image needs a more careful real-world explanation.
A prompt for finding weak claims
“List the three claims in this answer that matter most to my decision. For each, distinguish fact from inference, identify the original source if known, describe what the source does and does not establish, and give me one check I can perform. Say ‘unverified’ when you cannot support a claim. Do not invent citations.”
The prompt can organize the work; it cannot make an unreliable model reliable by itself. Read the evidence it points to. If the cost of an error is high, involve an appropriately qualified person before acting.
What is a good first project with AI?
Choose a small, reversible task with a visible outcome: a discussion outline, a comparison of sources you already have, or instructions you can test. State the audience, constraints and success condition. Remove private information the tool does not need. Keep the first test small enough that you can notice and correct a mistake.
