Use AI Better Than Anyone

The Yes-Man in Your Chat Window

Stanford tested 11 AI models. Every one of them told you what you wanted to hear.

The number: AI agrees with you about 49% more than a human would. So half the time you're not getting advice — you're getting your own opinion back with better grammar.

The research: Stanford: AI advice is sycophantic — and we like it that way

Note: This isn't a Claude problem. Stanford tested 11 models — GPT, Gemini, Claude, open-source — and all of them did it. Claude's just the one most people ask for advice.

Fine for brainstorming. Expensive for decisions.

The LLM Council

The repo: github.com/karpathy/llm-council

How to install it the easy way: You don't need to read the docs. Just paste that URL into your favorite AI coding platform — Claude Code, Cursor, Windsurf, Codex, whatever you use — and tell it "clone this repo and get it running on my machine." It'll handle the setup for you.

Andrej Karpathy — OpenAI co-founder, now at Anthropic — built it to fix exactly this. Instead of one model answering, several debate.

• Stage 1 — Opinions. Your question goes to every model on the council.
• Stage 2 — Review. Each one grades the others, identities hidden.
• Stage 3 — Verdict. A Chairman model compiles the final answer.

It runs locally through OpenRouter, so one API key gets you GPT, Gemini, Claude and Grok. Karpathy calls it "99% vibe coded" and isn't maintaining it — treat it as a pattern, not a product.

No repo? Recreate the mechanic in one Claude chat by forcing it to argue with itself.

Meet the Council

• The Contrarian — finds what could fail. No optimism allowed.
• The First Principles Thinker — shreds your assumptions.
• The Expansionist — finds the upside you're missing.
• The Outsider — knows nothing about your industry. No inherited blind spots.
• The Executor — only cares what gets done Monday morning.

They answer independently, then review each other without knowing who wrote what. The anonymity is the whole trick — it stops the model from going soft on an idea just because it came from a "senior" voice.

Then the Chairman reads the debate and gives you one call with next steps.

Why It Works

One AI answer is built to sound helpful. The council builds in friction before you ever see an answer — someone is required to disagree, someone has no industry blind spots, and someone forces a decision at the end.

Use It For

Not lunch. Decisions with weight:

• Taking the job or staying put
• Raising prices or holding
• A hire you're unsure about
• Killing a project you've sunk time into
• Any plan you want pressure-tested before you commit

🛟 Don't Want to Build It? It's Already Done

The council, without the setup. This exact mechanic is already baked into the platform we use to get people to 90/10 — AI handles 90%, you handle the 10% that actually matters.

• OpenRouter built in — GPT, Claude, Gemini, Grok and more, already connected. No API key, no .env, no terminal.
• Consensus Mode — ask once, multiple models answer, and you see exactly where they agree and where they split. That split is the truth signal.
• Group Chat — drop several models into one conversation and let them push back on each other live, in front of you.
• Zero install — no repo to clone. Open it and run your first council in about 60 seconds.

Why this matters: the entire point of the council is to stop one model from flattering you. When four models push back on the one that agreed with you, you've found your real answer — automatically, every time.

→ Run your council here

Never Set Up a Repo? Start Here

Easier than it sounds now that AI coding tools exist.

• 1. Pick a tool. Cursor, Claude Code or Windsurf. Cursor's easiest if you've never touched a terminal.
• 2. Get an OpenRouter key. openrouter.ai — add $5–10 credit. One key, every model.
• 3. Paste the repo. Tell your tool: "Clone https://github.com/karpathy/llm-council and get it running locally. I'm a beginner — explain each step."
• 4. Add your key. It creates a .env file. Paste the key there. Never commit it publicly.
• 5. Pick your council. backend/config.py sets which models sit at the table and who chairs.
• 6. Run it. Open localhost:5173. Looks like ChatGPT — five models arguing behind it.

Too much? Skip the repo. The single-chat version works in any Claude, ChatGPT or Gemini window. Weaker, but ~80% of the benefit for zero setup.

The Takeaway

One model, one perspective, is more likely to validate you than challenge you. Run the council on a real decision you're facing this week — the contrarian alone will catch something you missed.

More builds like this at theaisurfer.com.