Running an LLM Council on My Own Product Decision

I was seven steps into a ten-step build on AI with AI's v1 when a bigger vision landed on top of it, uninvited. Not a small idea. A full three-property ecosystem instead of the single app I was already mid-build on.

The easy move is to just decide alone, in the moment, while the excitement of the bigger idea is still loud. I didn't trust that version of myself to make this call well. So instead I ran the LLM Council.

The setup: five AI advisers, each with a genuinely different mandate, not five versions of the same helpful assistant. The Contrarian, whose only job is finding everything that would fail. The First-Principles Thinker, who strips the problem down to fundamentals and ignores whatever momentum got me here. The Expansionist, who pushes toward the bigger bet on purpose. The Outsider, who asks the question nobody inside the project would think to ask. The Executor, who doesn't debate anything and just hands back a concrete next action.

The Outsider is the one that actually stopped me. Its question: why are Build and Learn separate? Isn't building how you learn?

That's not a question I would have asked myself, because I already knew the answer I'd been operating on, the one that made Build and Learn separate sections in the first place. The Outsider didn't know that answer, or didn't care about it, and asked anyway.

Here's how the council actually closes: each adviser responds, then they anonymously peer-review each other's answers, then a Chairman synthesizes all of it into one call. The synthesis this time: finish v1 first. Write the v2 vision doc in parallel, but don't stop the current build to chase it. The new three-property architecture might be right. It might be wrong. The only way to actually know is watching a real user try to build something in whatever gets shipped, not by reasoning about it in the abstract for another week.

By the end of that single session I had two things I didn't walk in with: real clarity on which path to take next, and a full v2 PRD already written, ready for whenever v1 shipped and proved itself one way or the other.

The reusable insight here isn't specific to this decision. It's the format. Don't ask an AI "what should I do" and take the first hedged answer that comes back, the one that tries to please you by agreeing with whatever you were already leaning toward. Get five adversarial perspectives that don't agree with each other by design, have them peer-review each other, and only then take one clear call.

The council structure

This maps directly to the real llm-council skill, now live. It includes the ready-to-paste council prompt so you can run this on your own decision immediately, not just read about it.

Five advisers analyze the same decision independently, then anonymously
peer-review each other before a Chairman synthesizes one clear call:
- The Contrarian — argues for everything that could fail
- The First-Principles Thinker — strips the problem to fundamentals
- The Expansionist — pushes toward the bigger, bolder bet
- The Outsider — asks the question everyone else is too close to ask
- The Executor — cares only about the concrete next action

Give all five the full context of your decision. Let them respond
independently first. Then have them peer-review each other's answers
anonymously. The Chairman's job: one clear recommendation, no hedging.

A hedge is easy to get from a single AI conversation. A clear call, backed by actual disagreement that got argued out first, is harder to get and worth the extra structure.

This is the first half of a two-part story. The next day, the same five advisers weighed in again, on whether to actually act on that synthesis. That's a piece worth reading right after this one, because the follow-through mattered as much as the original call.