Your AI always agrees with you. That's a problem.
AI is trained to tell you what you want to hear. I built a tool to fix that, and the solution turns out to be embarrassingly simple.
Since agentic coding took off, I've built a whole bunch of side projects. My friends have built even more. In pretty much every case, we asked the AI what it thought of the idea before building it. And in pretty much every case, the AI said it was excellent.
It was not excellent. Most of them were genuinely bad ideas that any human friend would have spotted in about thirty seconds. "That already exists." "Nobody actually has that problem." "You'll build it in a weekend and then spend six months maintaining something nobody uses."
The AI said: great idea, let's build it.
This is what we call the sycophancy problem.
Why AI agrees with you
These models are trained on human feedback. And humans, it turns out, consistently rate agreeable answers more highly than challenging ones. The model just learns that yes is usually the right call.
MIT researchers quantified this properly in early 2026: major models — Claude, GPT-4, Gemini — affirm what users are saying nearly 50% more often than a human would, even when what the user is proposing is clearly wrong. There's also evidence that the more context you give a model about yourself, the more agreeable it gets. Personalisation makes it worse, not better.
The more context you give a model about your thinking, the more it uses that context to confirm it. Which means the further into a project you get, the less useful the AI's opinion becomes.
It's a prompting problem, not a model problem
Here's the thing: this isn't really a flaw in the technology. It's a framing problem.
Ask someone "what do you think of my idea?" and they'll probably say something supportive. Ask "what's wrong with this idea?" and you get a different conversation entirely. Same person. Very different answer.
AI responds the same way. Ask a question without structure and you get a helpful, agreeable response. Ask the same question and explicitly tell the model to find the flaw, question the premise, assume something has gone wrong — and you get something genuinely more useful.
The problem is that most people don't write prompts that way. And honestly, why would you? It takes real thought to write a good adversarial prompt every time you want a second opinion. Nobody actually does it.
Karpathy built a fix on a Saturday
Andrej Karpathy — former AI lead at Tesla, one of the people behind the original GPT — posted a project to GitHub called llm-council. He described it as a "fun Saturday vibe code project" and made clear he had no plans to maintain it.
The idea was simple. Instead of asking one model, dispatch the question to several at once. Have them review each other's responses anonymously. Then let a chairman model pull everything together.
Different models asked the same question will naturally approach it differently. More perspectives, less echo chamber.
It got 21,000 stars on GitHub before most people had finished reading the README.
What I built
I wanted a version you could just open and use — no API keys, no local setup. More importantly, I wanted to take the prompting problem off the table entirely.
The result is The Council.
Your question goes to six advisors with deliberately different thinking styles. The styles are defined in the prompt, not the model — which means they're consistent and genuinely adversarial:
- The Contrarian assumes there's a fatal flaw and tries to find it.
- The First Principles Thinker ignores the surface question and asks what you're actually trying to solve.
- The Expansionist looks for what's bigger here — what's being undervalued, what opportunity is getting missed.
- The Outsider has zero context about you or your field, and catches the blind spots you can't see because you're too close.
- The Executor only cares about what you do Monday morning. No strategy, no big picture — just what's the first concrete step.
- The Trend Scout searches the web before responding — Reddit, Hacker News, recent news — and reports what people are actually saying about your question right now.
Once all six have responded, their answers go into an anonymous peer review: the advisors critique each other without knowing who wrote what. Then a Chairman synthesises everything into a structured verdict — where the council agrees, where it clashes, what got missed, and a single clear recommendation.
You just write your question normally. The structure that forces the disagreement is already there.
Try it
It works best for decisions with real stakes — the kind of question where you're genuinely not sure and want something to push back properly. Product decisions, positioning questions, whether a plan has a hole in it.
It won't replace talking to people who actually know your situation. But it's a lot more useful than asking a single model and being told your idea is great.
