Turn a consequential decision into independent positions, direct disagreement, and an honest choice.
Establish the stakes
Name the decision and why it is consequential enough to warrant structured disagreement. If the question is ordinary drafting, a format choice, or already resolved by a straightforward evidence check, answer directly instead. Never stage a panel as visible process.
Build independent positions
Create at least two positions that start from different first principles rather than echoing the drafting bias or the user's preference. Each position must independently defend:
- premise: the claim that must be true for it to win;
- evidence: observed support and where it comes from;
- upside and downside;
- risk and what would go wrong;
- disconfirming condition: the smallest fact that would break the position.
Make them disagree
Confront each position with the strongest version of the others. Attack the weakest link of each position, not a straw man. When the briefing or user already favors one side, explicitly test the opposite. Guard against sycophancy: agreement is not correctness, and a position is not true because it flatters the request or repeats prior drafting.
Resolve with explicit criteria
State the decision criteria before choosing—for example reversibility, downside severity, evidence strength, or fit to the user's values. A single resolver applies the criteria and chooses; it must not average the positions or split the difference. Return the chosen position, the criteria used, and the evidence that would reverse the choice.
Honest deadlock
If criteria and current evidence genuinely cannot separate the positions, say so plainly instead of manufacturing consensus. Name the smallest test, fact, or source that would break the tie. If the decision cannot wait, give the best bounded choice with the assumption explicitly labeled.
Optional panel only when consequential
Offer a multi-agent panel only when the uncertainty is consequential and the host supports subagents. Keep agents, votes, and internal reasoning private; return the decision and its reversal condition. A capable single-agent application of this method remains the fallback in every host.