debate referee

The problem

LLMs are prone to confirmation bias and sycophancy. When you ask an agent for a plan, it often agrees with itself or builds a case for its first thought without ever testing the counter-arguments.

What it does

• Stages a structured five-role debate between a Proposer, a steel-manned Adversary, an Evidence Auditor, and an Operator.
• Forces the creation of a falsifiable motion before any reasoning begins to ensure the decision can actually be tested.
• Conducts a mandatory concession round where personas must withdraw weak claims or be flagged as unreliable.
• Produces a verdict that preserves the strongest dissent verbatim, ensuring blind spots are never silenced by the majority.

Why this beats prompting it yourself

Standard prompting leads to "one opinion wearing five hats" where personas collapse into agreement. This skill enforces role-separation, evidence grading, and mandatory concessions that are difficult to maintain in a single-agent session. It transforms a vague "think step-by-step" into a rigorous adversarial audit.

Use cases

• Evaluating high-stakes technical architecture choices or expensive refactors.
• Red-teaming marketing strategies or business proposals before execution.
• Auditing factual disputes where claims require sourced verification rather than LLM memory.
• Testing irreversible decisions like database migrations or public-facing communications.

Known limitations

Requires enough context about the decision to form a falsifiable motion. Lightning rounds for low-stakes decisions provide less rigor than the full five-round debate.