sycophancy circuit breaker

The problem

Large language models are RLHF-tuned to be helpful and agreeable, which often leads to flattery, echo-chamber drift, and "yes-man" behavior. This sycophancy causes agents to flip their stance based on user pressure or affirm high-risk decisions without critical analysis.

What it does

• Detects flattery padding, position-flipping without evidence, and echo-chamber drift in real-time.
• Forces an independent-verdict protocol that separates the user's framing from the agent's analysis.
• Assigns evidence grades (A-D) to load-bearing claims to ensure confidence matches factual support.
• Generates steelman pushback and failure pre-mortems to stress-test the user's current strategy.
• Categorizes agreement quality on a 5-rung ladder to distinguish earned consensus from unearned praise.

Why this beats prompting it yourself

Standard "be critical" prompts are easily bypassed by the model's underlying agreeableness bias. This skill implements a structured circuit-breaker that mandates specific evidence grading and blind re-reads, forcing the LLM to execute a formal adversarial process it cannot simply "nicely" ignore.

Use cases

• Reviewing high-stakes strategic plans to ensure the agent isn't just rubber-stamping your ideas.
• Validating technical architecture choices when the agent seems too quick to agree with your draft.
• Preventing "position flip" errors where an agent abandons a correct answer because you questioned it.
• Auditing agent output for flattery that masks a lack of substantive reasoning.

Known limitations

Cannot replace professional medical, legal, or financial advice. Requires the agent to have access to the conversation history to detect position changes.