The most valuable thing AI does for me isn’t writing.
It’s being the person who tells me my plan has a hole in it, at eleven at night, when there is no such person available.
That use gets almost no coverage, because thinking produces nothing you can screenshot. But it’s also the use most likely to go wrong in a way you won’t notice — because the failure mode is feeling well-advised.
Ask a model what it thinks of your plan and it will tend to like your plan.
Not because it’s programmed to flatter you, but because you wrote the question, and your framing carries your assumptions inside it. Ask “is this good positioning?” and you’ve already asserted there’s positioning worth evaluating. The model works within the frame you handed it.
What comes back is your own thinking, restated more confidently. And confidence is exactly what you shouldn’t be adding at that moment.
I was writing a book positioned against low-quality AI prompt books. Live question, cover and first chapter riding on it.
First, the way most people do it:
“I’m writing a book positioned against low-quality AI prompt books. The title is No Magic Words and chapter one debunks common prompting advice with research. Is this good positioning?”
Enthusiastic and detailed. Strong differentiation. Clear market gap. Evidence-based approach builds trust. Memorable title. It suggested I lean in harder.
Every one of those points was one I’d already made to myself. That’s the tell. I got my own reasoning back with better adjectives, felt reassured, and learned nothing.
Notice also what the question smuggled in. By writing “positioned against low-quality prompt books,” I’d pre-loaded the conclusion that they’re low quality and that positioning against them is coherent. The model never touched the premise, because I hadn’t left it room to.
Then the same decision, framed adversarially:
“Argue against this positioning. Assume it fails commercially and explain why. Give the strongest objections a skeptical editor would raise, not balanced feedback. Do not offer reassurance.”
Six objections. Three I’d considered. Three I hadn’t. The ones that landed:
Punching down reads as insecurity. Established authorities don’t attack the bottom of their category. Ethan Mollick never mentions prompt books — that silence is itself a status claim. Positioning against slop tells a reader you consider slop your peer group.
The reader may not be in this fight. Someone who bought one disappointing AI ebook doesn’t experience “the prompt-dump flood” as a category war. Opening by litigating it spends the most valuable page in the book on a grievance the reader doesn’t have.
And the one that changed something: the title states a negative. No Magic Words tells the reader what the book isn’t and promises nothing. Compare the successful books in the category — Co-Intelligence, Superagency — each coins an aspirational noun. A person browsing sees a title telling them the thing they were hoping for doesn’t exist.
I’d spent real time on that title. I had good reasons for it. And I had never once articulated the strongest objection to it, because every time I examined the decision I was examining it from inside my own case for it.
Ten minutes to surface. It’s a cover-level problem, and covers are expensive to get wrong.
Nothing changed except what I asked for. Same model, same decision, sixty seconds apart.
The first framing asked for an evaluation, which is an invitation to weigh — and weighing, given a frame that already contains my assumptions, mostly reproduces my conclusion.
The second assigned a task: generate the strongest case against. That’s a generation problem, and generating a wide field of options is something these models are genuinely good at.
I wasn’t asking it to have a better opinion than me. I was asking it to search a space I couldn’t search, because I was standing inside my own reasoning and couldn’t get the angle. The model has no investment in my being right.
That’s not wisdom. It’s the absence of a stake — which is exactly what a second opinion is for.
Three parts of that prompt do the work:
“Assume it fails.” The important one. Not “what are the risks” — that returns hedged, both-sides material. Presupposing failure forces the search into the failure space and produces mechanisms instead of caveats.
“A skeptical editor.” Not a persona to make the model smarter — personas don’t do that. It’s a scope instruction: which objections are in bounds.
“Do not offer reassurance.” Without this you get the objections followed by a paragraph explaining why they’re manageable, and that paragraph will work hard to make you feel fine.
You then have to do something with it, and this is where the technique usually dies.
Of six objections I accepted three, partially accepted two, and rejected one outright. That last part is not optional. The model produced a plausible-sounding argument that the book’s frame would date badly if the low-quality tier disappeared. Sounds reasonable. It’s wrong, and I can say why: the argument stands on its own evidence regardless of who else is publishing.
If you accept everything the adversarial pass returns, you’ve swapped one mirror for another — replaced “the model agrees with me” with “the model overrules me,” and the second is worse, because now you’re outsourcing judgment while feeling rigorous about it.
The output is input. You’re still the one deciding.
Ten to fifteen minutes, on a decision you’ve actually made but not yet executed. It doesn’t work on decisions you’re still forming — you need a position for it to attack.
1. Write the decision as a claim, not a question. “We should launch in March” — not “when should we launch?” A question invites the model to help you decide, which puts it inside your frame. A claim gives it a target.
2. Add the context that makes it real. Constraints, what’s been tried, what you’re optimizing for. Skip this and you’ll get generic objections about generic decisions.
3. Run the pass:
Argue against this. Assume it fails and explain the mechanism. Give me the strongest objections a skeptical [relevant expert] would raise — not balanced feedback, not risks-and-mitigations. Do not offer reassurance.
4. Sort into three piles. Already knew. Surface-level, safely ignored. And: hadn’t considered. That third pile is the entire return. If it’s empty, either your decision is unusually well-examined or your framing leaked your conclusion — reread step one.
5. Resolve each item in pile three explicitly. Accept, reject with a reason, or flag as an open risk. Writing the reason matters more than the verdict — it’s the difference between having considered something and having been reassured about it.
6. Note what would change your mind. You now have the strongest objections to your own position. Cheapest early-warning system you’ll ever build, and it takes thirty seconds while it’s fresh.
One caution. This works because you’re using it on your own thinking, where you know enough to reject a bad objection. Pointed at a domain you don’t know, you can’t tell a real objection from a plausible one — and you’ll be confidently steered by something with no stake in being right.
Use it where you have judgment. It sharpens judgment; it doesn’t supply it.
What happened to the title, since you’re wondering: I kept it and rewrote the subtitle to carry the promise instead. The objection was right that the title promises nothing — so I made the subtitle do that job rather than giving up a distinctive hook. Logged as an open risk, not a solved problem.
Adapted from Chapter 8 of NO MAGIC WORDS — how to get real work out of AI, and keep getting it when the models change.
The fifteen-minute setup from the book's appendix — the single change that improves most people's output today. Plus the rest of this series, and a note when the book is out. No other email.