There's a difference between asking AI "what should I do?" and asking AI "here's my lean, what am I missing?"
It's bigger than I expected.
Here's what I caught a few weeks ago. I'd describe a problem, list two or three options, ask the question, and get back something that read like a consultant deck. Summary of each option's pros and cons. A polite gesture toward "it depends on your priorities." Maybe a hedged recommendation. Useful for organizing my own thinking, but it rarely changed my answer. The AI was brainstorming with me, not pushing back.
The fix was small: I stopped ending my prompt with "what should I do?" and started ending it with "here's what I'm leaning toward and why. What am I missing?"
The output shape changed completely.
Saturday's example. I'm building a portal where field reps can review past sessions, and I needed an auth model. Three paths: passwords, magic links from the desktop app, or email magic links. I drafted a brief for two AI models with my lean already staked: Path B for V1, Path C in V2 once a rep asks for mobile. The reasoning behind the lean: preserves zero-password philosophy. The honest cost: doesn't solve mobile, which probably matters more than I'm weighting today.
What came back wasn't three more options. It was specific corrections.
One model flagged that my "one-time use" idea had a flaw I hadn't seen. Under certain conditions, the same login link could be redeemed twice. The other model questioned whether my "wait until a rep asks for mobile" trigger was the right gate, because the infrastructure I'd need for the future change had implications I should think through pre-launch. Both reads pointed at things I'd missed in my own brief. Neither would have surfaced if I'd asked "what should I do?"
Here's my theory on why the shape change matters. AI assistants default to helpful-executor mode. An open-ended "what should I do?" invites them to solve the problem from scratch, which means generating options. A staked lean is a target, and targets get critiqued. The same model can produce two completely different outputs depending on whether I asked it to brainstorm with me or react to me.
There's a second move I've been pairing with this. When I send the staked-lean brief to a second AI for an outside read, I prepend one line that primes the role: "You are not the executor. Review the brief and tell me what would change YOUR mind about the lean if you were the one making this call. Do not rewrite the proposal." Without that line, even with a lean staked, the second AI tends to slip back into helpful-rewriter mode and produce its own three-option brief instead of critiquing mine.
Two lines together: stake a lean, prime the reviewer.
The paradox is that AI feels more useful when I do more of the work upfront. Asking for opinions is faster to write, but the output is cheaper too. Drafting a position takes more time, but it's the difference between getting a critique that tightens the build and a summary that doesn't.
Try it on your next real decision. End your prompt with "here's what I'd lean toward and why, what am I missing?" See what shape comes back.
~James
This article first ran on LinkedIn. Comments are open there. Back to all writing.