A Claude skill that runs multiple sub-agents to roast your idea from different angles. Finally, a fix for sycophantic AI.
Single-agent AI feedback is broken by design. You ask Claude if your idea is good, Claude tells you it’s great. You ask for critique, Claude gives you a compliment sandwich. Andrej Karpathy’s LLM Council concept addresses this directly by spinning up multiple independent agents, each assigned a different perspective or persona, so they’re structurally forced to disagree with each other rather than with you.
Someone has now built that as a native Claude skill. You drop it into your Claude environment, give it an idea, and it fans out to distinct sub-agents that each critique the concept from a different angle. The output isn’t one hedged response. It’s a small council of reviewers.
What the Skill Actually Does
Each sub-agent gets a role. Think devil’s advocate, skeptic, domain expert, user advocate. They evaluate the same input independently, then the outputs get surfaced together. You see the tension between perspectives rather than a smoothed-over consensus.
This matters because the sycophancy problem in LLMs isn’t a vibe issue. It’s a training artifact. Models are RLHF-tuned on human approval signals, so they optimize for responses you’ll rate highly. Disagreement feels bad, so the model avoids it. Multi-agent architectures sidestep this by making each agent’s job to critique, not to please you.
How to Use This Right Now
Pull the skill into your Claude setup. The creator (@Hesamation is crediting a builder who’s been shipping multiple skills) has packaged it so it runs inside Claude’s tool/skill framework.
Use it on decisions that feel obvious to you. That’s exactly when you’re most blind. Feed it a business idea, a technical architecture decision, a product pitch, a feature spec. The more you already agree with yourself, the more you need this.
A few practical tips:
- Treat the harshest sub-agent as signal, not noise. Your instinct will be to discount it. That’s the sycophancy reflex showing up in you, not just the model.
- Run it before you’ve socialized the idea. Once you’ve pitched it to three people and they liked it, you’re already anchored.
- Look for which critiques cluster. If two or three different sub-agent personas land on the same problem from different directions, that’s a real issue.
The underlying pattern here is worth generalizing beyond this specific skill. Any time you’re using a single AI call to evaluate something, you’re getting one path through the probability distribution. Multiple agents with different priors get you something closer to a distribution of opinions. That’s more useful than one confident answer, especially for decisions with real stakes.