The method

What is the Council Method?

The Council Method is a system for getting answers out of AI you can actually trust, created by Jason Santiago and taught in his book Let the AI Be Smart. It runs in three passes. First, you ask several different AI models the same question separately, so none of them can copy another. Second, you show each model the other models’ answers and make it defend or correct its own, and every claim has to name a source you can open, because models agreeing is not proof. Third, one model you trust weighs all of it into a final verdict that keeps only what survived, and a person signs it. It solves one specific problem: a single AI sounds exactly as confident when it is wrong as when it is right, and on real work, confident and wrong gets expensive.

The Council Method is also called the Council Methodology.

Why one model is never enough

A single AI response is one witness's account. It may be excellent. It may be wrong. It will sound the same either way, because fluency is what these systems are best at. The book puts it plainly:

One response is not a verdict. It is testimony.

Let the AI Be Smart

You would never convict on one witness in a matter that counted. The Council Method applies that standard to AI answers that matter: business decisions, technical choices, anything you will have to stand behind later.

Why independent investigation comes first

In a Council run, every model answers blind, before seeing any other answer. This is not a detail. If models see each other's answers first, the first confident voice anchors the room, and you get an echo instead of an investigation. Independence first, conversation second.

Then you read the results by where the disagreements cluster. Divergence is not a nuisance. It is the signal. Where the models fork is exactly where the real question lives.

Why agreement is not verification

Consensus is cheap. Four models agreeing costs you nothing and proves you nothing.

Let the AI Be Smart, Chapter 8

Models trained on similar data can be wrong together. So the method demands provenance before consensus: every claim must name where it came from, and every claim must survive an attempt to kill it, before it earns a place in the final answer. Claims that cannot name a source get demoted. They can never sneak into the verdict dressed as facts.

The verdict stays human

The end of a Council run is not an average of the answers. It is a decision: what everyone established, what only some did, what remains unresolved and gets carried forward on the record. And a person signs it. The machine can make the answer survivable. Only you can make it yours.

Where it came from

Jason Santiago developed the method on production work, not in a lab. As a commercial construction superintendent, wrong answers carry schedule, money, and safety costs, and a confident guess is the most dangerous kind. The turning point was a real build where one model confidently invented a software component and could not see its own error, and a second model spotted it instantly. The lesson generalized: stop asking one machine to be right, and start building a room where being right has to survive the room.

In practice: how you actually run one

A Council run is three passes, and you can do the whole thing by hand in the chat windows you already have. No code, no tools you have to buy.

1. The independent pass. Open two or three different models in separate chats (five if the question is big). Paste the same question and the same context into each one. Do not let them see each other. Collect every answer.
2. The cross-examination pass.Now copy the answers and hand each model the other models' answers, never its own. So the answer from one model goes to all the others, not back to itself. Ask each one to challenge the weakest claim it sees in the room, then defend or revise its own answer based on what the others said. Collect the revised answers.
3. The synthesis pass. Take all of it, the independent answers and the revised ones, and paste it into whichever single model you trust most as your synthesizer. Ask it for a verdict that keeps only what survived, names where the room agreed and where it split, and states honest confidence. You read that verdict and you sign it.

The copying is the entire manual overhead, and it is exactly what we build platforms to automate. The Council Room on this site does all the passing and collecting for you so you can watch it happen. But you do not need any of that to run the method. You need a few chat tabs and the discipline to keep them honest. The full protocol, with copy-paste prompts for one, three, or five models, is a free tool: the Council Protocol.

Patent pending

The system behind the method is the subject of a patent application Jason Santiago filed on December 5, 2025: Autonomous Multi-Agent Code Generation with Mandatory Security Diff Validation and Instantaneous Atomic Rollback. The application covers the multi-model coordination the Council Method is built on: a supervisory agent, mandatory validation that rejects prohibited patterns, immutable version records, and single-action rollback.

The Council Method was developed by Jason Santiago and is taught in full in Let the AI Be Smart.