Judge Panel

Hard Rules

• Angles are load-bearing. Every parallel attempt gets the SAME task but a DIFFERENT strategy angle. If you give all attempts an identical prompt, you get N rephrasings of one idea — the agreement-collapse failure this skill exists to defeat.
• Judges must be independent of authors. An attempt's own author may never score it. A judge that shares an author's reasoning shares its blind spots.
• Score concrete, named dimensions — never "which is better?" Vague rubrics produce noise. Define 3–6 scorable criteria up front.
• Synthesize; do not just crown the top score. Graft the specific strengths judges flagged in the runners-up. Discarding them throws away half the value.
• Don't use this on narrow tasks. It burns N× the tokens. Reserve it for genuinely wide forks where a wrong early commit is expensive.

Judge Panel (parallel attempts → independent judging → synthesis)

A structured way to beat the "first idea, iterated" trap on wide-open problems. Instead of committing to one approach and polishing it, you generate several genuinely different attempts in parallel, judge them against explicit criteria, and synthesize from the winner — grafting the best ideas from the runners-up. Grounded in the multi-agent finding that diversity of attempt is the dominant driver of quality: heterogeneous attempts beat homogeneous ones, and correlated errors among same-prompt agents collapse the gains.

When this beats just doing it

• The solution space is wide and the best shape is genuinely unclear (data schema, API surface, algorithm choice, refactor strategy, page layout).
• A wrong early commit is expensive to undo.
• You can state what "good" looks like as concrete, judgeable criteria.

If none of those hold, skip this — generating several attempts at an obvious task is waste.

Step 1 — Define the contest (do this in your main working context, before spawning anything)

1. The task: one precise sentence of what each attempt must produce.
2. The angles: name 3–5 genuinely different strategies to seed diversity (e.g. "MVP-first", "performance-first", "maintainability-first", "use-an-established-library", "from-scratch"). Same task to all + a different angle each — that's what breaks agreement-collapse. Homogeneous attempts defeat the purpose.
3. The judging criteria: 3–6 concrete, scorable dimensions (correctness, simplicity, extensibility, performance, test surface). Vague criteria → vague judging.

Step 2 — Run the tournament

Three phases: Attempts → Judge → Synthesize.

Attempts (parallel). Spawn one sub-agent per angle. Each gets the identical task plus its assigned strategy:

Produce an attempt at: <TASK>. Use the "<ANGLE>" strategy specifically. Return the concrete artifact plus a short rationale.

If your host has a parallel-agent or workflow primitive, fan these out concurrently. If it doesn't, run them sequentially in fresh contexts — the key invariant is that each attempt is produced independently, without seeing the others (so they don't converge).

Judge (parallel, independent). Each completed attempt is scored by a separate agent against the same rubric — independence matters more than redundancy:

Score this attempt against these criteria: <CRITERIA>. Be specific and critical. Return a per-criterion score, a total, and the single strongest and weakest aspect.

ATTEMPT: <the artifact>

Have each judge return structured output so the results are sortable, e.g.:

{
  "scores": { "<criterion>": 0 },
  "total": 0,
  "strongest": "...",
  "weakest": "..."
}

Rank. Sort the judged attempts by total score, descending. Carry every judge's strongest/weakest notes forward into synthesis — they are the cross-pollination signal.

Step 3 — Synthesize (main context)

Read the ranked results. Do NOT just take the top score blindly:

• Adopt the winner as the base.
• Graft specific strengths the judges flagged in runners-up (the panel's value is cross-pollination, not just ranking).
• If the top two are within a hair, name the tradeoff explicitly and pick with a one-line justification.
• Implement the synthesized result in your main context, then verify normally (tests / run).

Trigger Conditions

• "What's the best way to structure X" where X has many viable shapes.
• "Try a few approaches to Y and pick the best."
• "I'm not sure which design wins" / "compare a few implementations."
• Competing-implementation or competing-design forks where the right answer isn't obvious up front and a wrong early commit is costly.

Out of Scope

• Narrow implementation with one obvious approach → just implement it (optionally plan it first with spec-driven-dev).
• Adversarial stress-test of a single proposal → devil-advocate.
• A focused review of code already written → code-review-session.
• Verifying one discrete factual claim → that's fact-checking, not a design tournament.

Common Traps

• Homogeneous attempts — if all agents get the same prompt with no distinct angle, you get N phrasings of one idea (the agreement-collapse failure mode). The angles are load-bearing.
• Self-judging — never let an attempt's own author score it; judges must be independent. A model judging its own work shares its blind spots (the correlated-error failure).
• Vague rubric — "which is better?" yields noise. Score concrete, named dimensions.
• Over-use — this burns N× tokens. Reserve for genuinely wide forks; for narrow tasks it's pure overhead.
• Skipping synthesis — taking the raw winner and discarding the runners-up throws away half the value. Graft.
• Attempts seeing each other — if the parallel attempts can read one another mid-generation, they converge and you lose the diversity you paid for. Keep them isolated until judging.

Skill Chain