Consensus

Consensus

Structured decision evaluation through stance rotation — analyze from advocate, critic, and neutral perspectives, then synthesize into a confidence-rated recommendation with concrete next steps.

Workflow

Step 1: Select Mode

| Decision Scope | Mode | Reason | | -------------- | ---- | ------ | | Bounded, reversible | Sequential (default) | All perspectives in one pass — fast | | Multi-month or irreversible | Subagent | Three isolated subagents prevent cross-contamination | | Perspectives suspiciously aligned | Escalate to Subagent | Lack of genuine disagreement signals contamination |

Use subagent mode when: the decision impacts more than 3 months of work, multiple teams are affected, or sequential perspectives align too easily (suspiciously low disagreement likely signals contamination — isolated subagents are required to get genuine divergence).

See references/consensus-strategy.md for full escalation criteria.

Step 2: Stance Rotation

Rotate through three perspectives (see references/stance-rotation.md for detailed prompts):

1. Neutral — state the decision, list all factors (technical, organizational, timeline, risk), note missing information, present assessment without leaning toward a conclusion.
2. Advocate — build the strongest case FOR: what problems does it solve, what synergies does it create, how can challenges be overcome? Subject to ethical guardrails — refuse to advocate if fundamentally harmful.
3. Critic — rigorous scrutiny: real risks, overlooked complexities, simpler alternatives, flawed assumptions? Subject to ethical guardrails — acknowledge if the proposal is genuinely sound.

In subagent mode, dispatch three isolated subagents (one per stance) with identical context. Subagents must NOT see each other's output.

Step 3: Synthesize

Weigh all three perspectives and produce a recommendation:

1. Points of agreement — where all perspectives align (strong signal)
2. Points of disagreement — where they diverge and why
3. Recommendation — with confidence level: low / medium / high
4. Would change if — conditions that would flip the recommendation
5. Next steps — concrete actions based on the recommendation

Validation Checkpoint

Before delivering the synthesis, verify:

• [ ] Each perspective contributed at least one unique point not raised by the others
• [ ] The critic identified at least one genuine risk (not manufactured disagreement)
• [ ] The recommendation confidence level is justified by the degree of inter-perspective agreement
• [ ] If all three perspectives agree too easily, escalate to subagent mode

Example

Decision: "Should we migrate from REST to GraphQL?"

| Perspective | Key Finding | |-------------|-------------| | Neutral | Current REST API has 47 endpoints; clients use ad-hoc field filtering. GraphQL would reduce over-fetching but adds schema maintenance. | | Advocate | Mobile clients would cut payload size ~60%. Single endpoint simplifies versioning. Strong ecosystem tooling available. | | Critic | Team has no GraphQL experience — 2-3 month learning curve. Caching is harder. Existing REST clients need migration path. |

Synthesis:

• Agreement: Current API has over-fetching problems worth solving.
• Disagreement: Whether the learning curve cost is justified given timeline.
• Recommendation: Adopt GraphQL for new endpoints only (confidence: medium).
• Would change if: Team had prior GraphQL experience (→ high confidence, full migration) or deadline is

References

• [Consensus Strategy](references/consensus-strategy.md) — Mode selection and escalation criteria
• [Stance Rotation](references/stance-rotation.md) — Detailed rotation steps, subagent dispatch, synthesis framework
• [Stance Prompts](../perspectives/references/stances.md) — Advocate, critic, neutral prompts with ethical guardrails (from perspectives skill)

Guardrails

Add guardrails instructions here.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: cofin
• Source: cofin/flow
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.