best-answer

Best Answer

Build the single strongest answer to one open question: a diverse panel attacks it, a clean-room judge maps their disagreement, then you synthesize one grounded answer. The engine is diversity of error — independent attempts fail differently, so you resolve the differences deliberately instead of averaging them.

When to use

You have a hard question and no answer yet, and you want the single strongest one.

• NOT this if you already have an answer and want it checked → validate-answer.
• NOT this if you want a list of problems in a draft, not an answer → find-gaps.
• NOT this if you're grading one concrete proposal → second-opinion.

Instructions

Step 1 — Frame the question and design the panel

Restate the question in a line or two. Pick a panel size (default 3; range 2–4).

Give each panelist a distinct mandate — angles that attack the question from non-overlapping directions:

Task type	Example panel
Research / analysis	first-principles mechanism · skeptic (what's overstated) · practitioner (how it plays out)
Diagnosis	each panelist pursues a different culprit class
Design choice	each authors a different approach, then argues its failure modes

When all panelists share one model (e.g. Opus), diversity must be manufactured from the mandates — identical panelists co-vary and the judge's map comes back empty.

Step 2 — Fan out the panel (parallel)

Dispatch all panelists in one message (multiple Agent calls) so they run concurrently. Each gets the same question, its own mandate, and tools (web search, bash, read). Ask each for a structured answer — its position, key claims with evidence, and a confidence — not a raw dump.

Step 3 — Map the disagreement (clean room)

Dispatch one general-purpose subagent as a clean-room judge — give it only the panel outputs (verbatim, each labeled by panelist) and the question, never your reasoning or which panelist you trust. Instruct it to:

• map the answers into three buckets — contradictions (incompatible positions, with each side), blind_spots (relevant points NO panelist raised), unique_insights (a point only one panelist made);
• earn each item — only real contradictions and material omissions, no padding;
• map, don't decide — pick no winner, write no final answer;
• surface, don't verify — don't run code or chase ground truth (that's Step 4);
• not spawn its own subagents;
• return the DisagreementMap verbatim — the structured map is the whole response, no narrative.

Step 4 — Verify and synthesize (you)

You hold the original context, so you write the answer:

• Resolve each load-bearing contradiction deliberately — check ground truth where it's cheap (read the code, run the check), dispatching a subagent per fork when needed. Don't split the difference.
• Fold in unique_insights that survive scrutiny.
• Address blind_spots, or flag them as open.

Write one grounded answer, attributing each load-bearing claim to its evidence.

Rules

• Recursion guard. Panelists must not fan out their own panels — keep deliberation to one level.
• Cost discipline. Reserve for hard, open-ended questions; answer a cheap or deterministic one directly.

---

Output Schema

DisagreementMap {
  contradictions: [ { topic: string, positions: string[] } ],
  blind_spots: string[],          // each: the gap + why it matters, one line
  unique_insights: [ { insight: string, raised_by: string } ]
}