Genuine minority dissent makes a group reason better: a person who truly holds a contrary view makes the majority search more broadly and consider more options, even when the dissenter is wrong. The catch, established by the same research, is that role-played devil's advocacy does not replicate this - assigned dissent gets discounted as performance. So an AI cannot be the dissent; anything a model argues against a plan is constructed, the weaker kind. This skill therefore does not pretend to be the dissenter. It engineers the conditions for real dissent: it audits whether genuine dissent exists, surfaces who holds it, plans how to elicit and protect it, and flags constructed dissent as constructed. The output is a dissent audit and plan.
When to Use
- A group decision shows suspiciously smooth consensus and nobody is really pushing back.
- You can influence how challenge is gathered (who speaks, anonymous input, outside reviewers).
- Before a high-stakes call where you want genuine, not performed, challenge.
When NOT to Use
- As a source of dissent itself: the model's contrarian view is constructed, not authentic (use
red-team-lightfor that, which is honest about being constructed). - In a purely solo setting with no access to other people - you cannot manufacture authentic dissent.
- When genuine dissent already exists and is being heard.
- To "assign a devil's advocate" and consider the job done (the evidence says that does not deliver the benefit).
Instructions
When asked to set up or audit dissent, follow these steps:
- Audit the consensus. Is the agreement genuine, or is it smoothness from anchoring, hierarchy, or conformity? Note signs (no one names a downside, the senior view landed first, dissent would be costly).
- Locate real dissent. Identify whether anyone actually holds a minority view, and whether it is being voiced, ignored, or suppressed.
- Label what is in play. Mark any current "dissent" as authentic (a real holder) or constructed (assigned/role-played/AI). Do not let constructed dissent count as the real thing.
- Plan to elicit and protect genuine dissent. Concrete moves: anonymous pre-reads, asking the quietest person first, bringing in an outside reviewer who genuinely disagrees, separating generation from evaluation, protecting the dissenter from cost.
- For high stakes, prescribe a real dissenter. Recommend finding a person who actually holds the contrary view, rather than relying on a constructed critique.
- Emit the dissent audit and plan per
references/TEMPLATE.md.
Output Format
Use the template in references/TEMPLATE.md. The deliverable is the audit plus an elicit-and-protect plan, not a constructed counter-argument (that is red-team-light's job).
Quality Checklist
Before finalizing, verify:
- The consensus is assessed for whether it is genuine or smoothed by hierarchy/conformity.
- Any dissent currently in play is labeled authentic vs constructed.
- The plan gives concrete ways to elicit and protect real dissent.
- For high stakes, it prescribes finding a real dissenter, not relying on constructed critique.
- The skill does not present the model's own contrarian view as authentic dissent.
- The output is the dissent audit/plan artifact, not prose.
Evidence
Tier S. Authentic minority dissent reliably broadens a group's thinking (Nemeth et al. 2001; In Defense of Troublemakers), and the same research shows role-played devil's advocacy does not replicate it. That negative result is load-bearing here: an AI's contrarian output is constructed, so this skill works on the conditions for real dissent rather than claiming to supply it. The evidence is for human groups; it bounds, not just transfers to, AI use. Full grading: evidence/dossier.md.
Examples
See references/EXAMPLE.md for a completed dissent audit and plan.