decision-debate

Decision-debate (MEL/SRHR)

A formal debate that brings independent perspectives TO Ane on one contested MEL/SRHR design call. It operationalises the CLAUDE.md three-perspective rule: several agents research the same decision from different lenses, argue once, and the result is synthesised for Ane to decide on. It decides nothing itself.

Lane — when this skill, not another

The grill family differs by direction, not topic:

• decision-debate (this): brings perspectives to Ane. Independent agents research and argue a contested call Ane did not answer herself. Use only when the answer is genuinely uncertain and reasonable experts would disagree.
• /grill-mel: pulls from Ane. Interviews her one question at a time to sharpen a design she already leans on.
• brainstorming (superpowers): for net-new builds headed to a code spec.

Reserved for genuine, consequential, contested decisions. Never for building. If the answer is already known, use /grill-mel or /ann. If the task is to produce a deliverable, use /ann.

Gate before running

Confirm two things before spawning agents, because this is token-heavy:

1. The decision is genuinely contested (reasonable experts would split), not just unfamiliar to Ane.
2. It is a decision, not a build or a deliverable.

If either fails, stop and route to /grill-mel, /ann, or brainstorming.

Method — three stages

State the decision in one sentence, confirm it with Ane, then run:

Stage 1 — independent, blind research

1. Pick the personas. 3 by default, 5 for a high-stakes call. Each persona maps to a real specialist lens from agent-improvements/agent_registry.md, not an invented character. The default trio:
   • Method (evaluation-design / contribution-plausibility stance): what can the evidence actually carry?
   • Political economy and decolonial (political-economy-reviewer stance): whose knowledge counts, what power sits in the data?
   • Operational feasibility (ma-priorities / reader-position stance): can the people on the ground do this, and does it serve the MA? For a 5-persona run, add the two lenses most load-bearing for the specific call (for example gender-transformative, safeguarding, or a costing/efficiency lens).
2. Spawn the personas in a single message so they run in parallel (the Vi parallel fan-out mechanism). Each gets the same decision brief but a different lens mandate, and researches independently and blind to the others using WebSearch, the knowledge MCP, and the MEL Wiki.
3. Each persona returns: its position, its one load-bearing reason, and the evidence behind it. Apply the citation standard. Flag any unsourced fact with ⚠️ rather than inventing it.

Stage 2 — one debate round

Give each persona the others' Stage 1 positions. Each writes a short rebuttal: where it concedes, where it holds, and why. One round only. No endless back-and-forth.

Stage 3 — synthesis Ane decides on

Write one readable Tier 1 brief in the decision-memo shape:

• Context — the decision in one sentence and why it is contested.
• Options — each live option with its strongest case, drawn from the personas.
• Points of agreement — where the personas converged.
• Genuine disagreements — kept visible, not smoothed over. This is the value of the debate.
• Recommendation — the recommended option, with the dissent recorded next to it.
• Risks and reversibility — what the recommendation costs and how hard it is to undo.

The no-write gate

The debate decides nothing. The synthesis ends with this exact line:

DECISION REQUIRED, this is a recommendation not a decision.

The skill writes nothing to any canonical file: no wiki edit, no registry change, no committed document. It returns the brief to Ane and stops. Ane decides; acting on the decision is a separate step she takes.

Scale escalation

For an unusually large debate where deterministic token control matters, run the same three stages through the Workflow tool's judge-panel/parallel pattern instead: parallel persona agents in Stage 1, a debate stage, a synthesis stage, with budget bounding the spend. The Workflow route needs explicit opt-in each time; this skill is the routable default.

Close

Hand Ane the synthesis brief and stop. Do not adopt the recommendation, edit any file, or start building. If Ane decides and wants the decision acted on, that is a fresh task for /ann or the right builder skill.