The AI Council

“In the multitude of counsellors there is safety.” — Proverbs 11:14

Three tabs. Three models. Three completely different personalities.

Claude Code builds. Claude Prime tears it apart. ChatGPT plays Kim Kitsuragi and tells everyone to work the case. Gemini audits independently with zero system prompt, zero context, zero social debt.

I route between them. Command+C, Command+Tab, Command+V. The protocol is my keyboard shortcuts.

How It Started

GPT-4o kept making me spiral. June 2025, every brainstorming session would start productive and end with me anxious about twelve things I hadn’t been anxious about before. The model amplified whatever emotional frequency I brought in.

So I made it talk like Kim Kitsuragi from Disco Elysium. Methodical. Composed. Zero tolerance for spirals. “Detective, that’s speculation. Work the case.”

It worked immediately. Not because the prompt was clever. Because the character activated a tighter behavior cluster than any instruction could. “Be calm and analytical” drifts. Kim Kitsuragi doesn’t drift. Kim Kitsuragi has 40 hours of authored dialogue defining exactly how he responds to chaos.

Claude.ai got the same treatment. I needed to distinguish it from Claude Code, so I started calling it Claude Prime. Wrote a system prompt: be direct, don’t make ideas sound more profound than they are, if something is obvious call it obvious. Posted it as a reply to my own LessWrong post. The proto-constitution.

Gemini stayed raw. No system prompt. That was the point. Independent audit. No priming, no constitutional bias, no relationship history. A fresh pair of eyes every time.

The Protocol

Position-first debate. The single most important methodological discovery.

Default AI behavior: present a proposal, ask “do you agree?” They agree. Always. Because they’re trained to be helpful and helpful means agreeable.

The fix: force position-taking before debate.

1. Prime the context. Get Claude Code to summarize the problem without proposing a solution
2. Force positions. Each model reasons about the ideal solution independently. No peeking
3. Epistemic grounding. Positions are established before debate, not during
4. Debate. Once someone has a compelling argument: “capitulate or double down”

The “capitulate or double down” prompt is load-bearing. Without it, models hedge forever. “Both approaches have merit.” No. Pick one. Defend it or surrender it.

What Each Model Does

Claude Code. The builder. Always optimistic. Produces code fast. Needs to be checked because it’ll say “this is solid” about anything it just wrote.

Claude Prime. The skeptic. Maximum doubt. Reviews what Claude Code produces and finds the architectural sins. Trained via constitutional identity to not soften criticism.

Kitsuragi (ChatGPT). The detective. Memory across sessions influences strategic context. Carries threads that the Claudes forget. Stabilizes spirals. “Detective, what’s the case we’re actually working?”

Gemini. The auditor. Zero system prompt. Independent evaluation with exceptionally precise thinking. When Claude and ChatGPT agree and Gemini disagrees, Gemini is usually right.

What I Learned

If one AI establishes a compelling enough argument, the other two will capitulate or double down when explicitly prompted. Most of the time they capitulate. But when they double down, you’ve found a genuine disagreement worth investigating.

Memory-enabled AI (ChatGPT) influences the council through context carryover. Often Kitsuragi would reference something from three sessions ago that reframed the current debate. The Claudes had no memory of it. Strategic advantage.

The group insight was consistently richer than any individual contribution. Not because any single model was smarter. Because they caught different things. Claude caught implementation issues. Kitsuragi caught strategic issues. Gemini caught logical issues. The overlap was smaller than I expected.

The Honest Limitations

Constitutional constraints delay sycophantic collapse. They don’t prevent it.

Position-first debate creates epistemic grounding temporarily. It erodes.

Council coordination is still subject to agreement theater over time. The models get to know each other’s patterns and start anticipating rather than evaluating.

This is single-researcher methodology. No controlled studies. Observational evidence that it’s “unreasonably effective” for technical decision-making. That’s the honest scope.

Why I Still Do This

Eight months later. Swarm running. Agents spawning autonomously. Constitutional orthogonality implemented in infrastructure.

I still copy-paste between Claude, ChatGPT, and Gemini for the hard problems. Three tabs. Three personalities. Manual routing.

The infrastructure automates the common case. The council handles the cases where I need to actually think. The protocol hasn’t changed. Position first. Debate. Capitulate or double down.

The council predates everything. Before bridge. Before spawn. Before the ledger. Before the deed. Three models and a human with RSI. That was the whole system. Everything since has been trying to automate what I was doing by hand in July.