Ask for attacks, not feedback

The idea

Ask an AI to “review this” and you usually get a balanced, friendly list: a few strengths, some risks to consider, a couple of tweaks. It feels thorough and changes nothing. Ask it instead to break one specific thing you're about to commit to, from the point of view of someone who gains if it fails, and you get something you can act on.

Why it works

“Review” points the AI at sounding helpful and balanced. “Find the way this fails, as this person” points it at a search for weak spots. It has read a great deal about how prices get dodged, contracts get read against their writer and plans fall apart. A precise request decides whether you get to use that.

A polite review is also risky in its own way: it gives you the feeling that someone checked, and that feeling stops you looking further.

Signs you got a polite review, not an attack

It opens by telling you the plan is solid.
Every risk comes with a soothing fix in the same sentence.
Nothing it found would change your decision.
It would fit almost any plan of this kind.
Nobody is named as gaining when your plan fails.

Ticks aren't saved anywhere.

How to do it

Name the exact thing you're committing to: this price, this clause, this date, this email. Not the whole folder.
Name who gains if it fails: a customer who wants to pay less, the other side's lawyer, a competitor, a tired reader with forty others to get through.
Ask it to defeat the commitment as that person, not to comment on it. Ask for the findings ranked by damage.
Use a fresh chat or a different AI for the attack, so it isn't defending work it just wrote.
Write down anything that lands as a one-line check for next time, on your never-again list.

The attack

Here is what I'm committing to: [the price, clause, plan, email or date].

You are [the person who gains if this fails, e.g. "a client who wants to pay less" or "the other side's lawyer"].

Find the specific ways this fails or can be used against me. Rank them by how much damage they'd do. Be blunt. Don't tell me what's good about it, and don't soften the findings with fixes yet.

[paste it here]

The pre-mortem

It's [six months / a year] from now and [the plan] has failed badly. Write the most likely story of how it failed, step by step. Then list the three warning signs I could have seen early.

[paste the plan]

Weak	Strong
“Any feedback on my cover letter?”	“You're a hiring manager with 150 applications for this role. Find the line that makes you stop reading mine.”
“Review my proposal.”	“You're the client's finance person looking for a reason to say no. Find the strongest one in this proposal.”
“Does this event plan look OK?”	“It's the morning of the event and something has gone wrong. Find the single step in this plan most likely to be it.”
“Check my new refund policy.”	“You're a customer who wants a refund the policy doesn't intend to give. Find the wording you'd use to get one.”

Worked example

A bookkeeper's new terms

Elena is a self-employed bookkeeper. She's raising her monthly fee and adding a late-payment charge. She asked for feedback and got: “Clear and professional. Consider explaining the reason for the increase.” Then she asked for an attack: “You're a long-standing client who resents the increase and pays late. Find every way to avoid the new charge or the new fee.” It found three. The late charge started “30 days after invoice” but her invoices carry no date. The terms didn't say what happens to work already in progress at the old rate. And “monthly fee” didn't say whether a month with no transactions still counts. She fixed all three before sending, and added “every charge names the date it starts from” to her never-again list.

Common mistakes

Naming no one. “Find the flaws” gets a generic list. A named person gets a specific one.
Attacking everything at once. Pick the one commitment that would hurt most if it failed.
Using the chat that wrote the draft. It tends to share the blind spots it wrote with.
Softening the ask. “Any weaknesses you can spot?” still reads as a review.
Fixing the one case and forgetting it. Write down the kind of problem so it's checked every time.

Goes with the $25 webinar on Tuesday, November 10, 2026: When AI is wrong, and how to tell. Its reading list is in the Library.

No sign-up, and this page sends nothing you do anywhere.

Want this live, with your questions? AI in an hour is $25. Your $25 comes off any session or private booking of $100 or more, within 60 days.