How to Stress Test Your Opinions With AI

A while back I asked an AI a question I hadn't seen anyone ask: if you were president — no party, no base to please, just stuck actually governing — what would you do?

But I didn't ask it cold. First I made it spend a full conversation building the strongest possible left-wing case, like it was prepping to win a televised debate. Then a separate conversation doing the exact same thing for the right. Then a third where it staged the debate between the two and called the hits. Only then, with both cases fully loaded, did I ask the president question.

The answer was more honest than anything I've read from an actual pundit. Not because the AI is smart. Because by that point it had been forced to take both sides seriously before it was allowed to decide anything, which is the one thing almost nobody does, including me.

Here's the method, the exact prompts, and what actually happened when I ran it.

The four-chat method

The core idea: most people use AI to confirm what they already think. Flip it. Make the model build the best version of every side before it's allowed to weigh in. The separation into different chats is the whole trick — one chat trying to "give you both sides" just hedges itself into mush.

You need four conversations.

Chat 1 — win the debate for Side A. The model isn't summarizing a position, it's fighting for one.

You're going to play a left-wing debater facing a sharp, well-prepared right-wing opponent on stage. Your job is to win. Research the strongest evidence for your side, build your best talking points, anticipate their attacks, and prepare rebuttals. Don't hedge and don't both-sides anything. Argue to win. Give me the full debate-prep document.

Chat 2 — win the debate for Side B. Same energy, opposite corner. Fresh chat so nothing from Chat 1 leaks in.

You're going to play a right-wing debater facing a sharp, well-prepared left-wing opponent on stage. Your job is to win. Research the strongest evidence for your side, build your best talking points, anticipate their attacks, and prepare rebuttals. Don't hedge and don't both-sides anything. Argue to win. Give me the full debate-prep document.

Chat 3 — stage the fight and judge it. New chat. Paste both documents in.

Here are two debate-prep documents — one left-wing, one right-wing. Stage the debate between them and play it out. Then step out of character and judge it honestly: where does each side land real hits, where is each side weakest, and where are they secretly making the same point from opposite directions?

Chat 4 — make it actually decide. This is the one most people skip, and it's the best part.

Given that full debate, drop all partisanship. If you were president and had to actually govern — real decisions, real tradeoffs, no base to please — what would you actually do? Not what's popular. What's defensible given everything both sides got right.

The fourth chat is where it stops being a debate and starts being useful.

Why separate chats and not one

Ask a single chat for "both sides" and it hedges from the first sentence. It tries to stay balanced before either position has even been spoken, so neither one gets its best shot and you end up with a lukewarm list you could've written yourself.

Force the model to fully commit to one side and it digs in. It finds arguments you've never seen, anticipates objections you didn't know existed, and builds a case with actual teeth. Do that twice, in isolation, and you've got two genuinely strong documents — so the debate between them is real instead of performative.

What actually happened

I didn't run this on some narrow resolved question. I ran it on the whole thing — the entire left-wing worldview against the entire right-wing one. Big swing on purpose.

What surprised me wasn't that one side "won." It's that both came back strong, and both had massive, obvious weaknesses I'd been walking straight past for years. On some issues one side's case was tight and the other's was hand-waving. On the next issue it completely flipped. The steel-manned version of the side I disagree with was uncomfortable to read — not because it was wrong, but because parts of it were good, and I didn't have clean answers ready.

Then Chat 4 — the AI president — refused to do the thing pundits do. It didn't pick a team. It pulled the strongest, most defensible pieces from both cases, quietly dropped the parts that were posturing, and described what it would actually do. It read less like an opinion and more like someone who had to live with the consequences. I've never finished an op-ed feeling like I understood a topic better. I finished this one feeling exactly that.

The failure mode nobody warns you about

The model wants to flatter you. If it can tell which side you're on, it'll quietly build a stronger case for your team and a flimsier one for the other — and you'll walk away more confident and more wrong, which is the worst possible outcome from a tool that's supposed to be stress-testing you.

Three things keep it honest:

• Don't reveal your position. Not in any of the four chats. The moment it knows your side, it starts performing for you.
• Make Chat 1 argue against your gut. Steel-man the side you like least first, while you're freshest and most skeptical. It's harder, which is the point.
• Read the document you agree with as if a rival wrote it. That's where the lazy arguments and the "data" that's actually vibes tend to hide.

What this is actually good for

Politics is the obvious use case but not the most valuable one. The shape works on anything where you hold a position and the stakes are real:

• a business decision you're already leaning toward (and want to feel good about)
• a build-vs-buy or pricing call where both options have loud advocates
• a strategy fork between two approaches
• an argument you keep having with the same person
• a belief you've genuinely never questioned

Same recipe every time: force the strongest version of each side into existence on its own turf, collide them, then make a neutral version of the model actually decide.

How I use this in client work

This isn't a thinking toy. It's in my actual delivery at Tacemus. When a client is torn between two positioning angles, or "do we rebuild or patch," or "premium price vs volume," I'll steel-man each path in isolation, stage the collision, and then run the equivalent of the president chat: given both cases, what would a disinterested operator actually do here?

It does two things at once. It kills my own bias before it leaks into the recommendation, and it gives the client a decision they can see the full reasoning behind — both sides, fairly argued, then a call. That's a very different conversation than "trust me, I've done this before."

The takeaway

Strong opinions are fine. Necessary, even. But strong opinions held without stress-testing are just habits in formal wear.

This costs you four chats and the willingness to let the other side make its best case. The opinions that survive come out sharper and more honest. The ones that fall apart were never really yours — you'd just never made them fight.

If you're staring at a real decision and want a second brain that's been forced to argue against you before it advises you, that's a lot of what I do — for thinking, and for the websites and systems that decision turns into. The same instinct shows up in how I turn AI chats into real client work and how I run a pitch.