build-tournament

name	build-tournament
description	Compete-mode for BUILT artifacts — landing page, component, small app, email template. Type /build-tournament <brief> to have isolated persona contestants each build the full artifact, a blind judge score them, and a revised champion ship. Reserve for high-stakes builds; one-shot first for routine work.
argument-hint	<the build brief — artifact, audience, constraints, success criteria>
disable-model-invocation	true
allowed-tools	Read, Grep, Glob, Bash, Write, Edit, Agent, Workflow, Skill

build-tournament — isolated contestants, blind judge, revised champion

Angela Jiang's (Anthropic head of product) compete mode: N isolated sub-agents each build the COMPLETE artifact from the same brief, a blind judge scores them against pre-declared criteria, and the champion is revised with the judge's feedback plus grafts from the losers. The revision round is where the quality comes from — raw contestant scores cluster within a few points.

When to invoke

• A high-stakes built artifact (landing page, dashboard screen, component, small app, branded template) where direction AND execution are both open and picking wrong is costly.
• You are about to one-shot a client-facing or launch-critical build — stop and run this instead.
• NOT for routine builds: a bare one-shot with the playbook in context is the baseline; this costs ~4x tokens and wall-clock, so the artifact must earn it.

Core procedure — DO NOT DEVIATE

1. Lock the brief and criteria BEFORE spawning. Write the full business brief (product, audience, CTA, constraints) and 4-6 scoring criteria with weights. Criteria declared after seeing output is judge contamination. UI artifact → run mobbin-ui-patterns first and put the reference brief inside the contestant brief (founder directive 2026-07-13).
   • Done when: brief + weighted rubric exist in a file contestants and judge both receive.
2. Spawn 3 isolated contestants via the Workflow tool (this skill's instruction satisfies its opt-in gate), one persona each, identical brief, isolation: 'worktree' when they write files, and explicit rules: complete the whole artifact, never ask for direction, never look at sibling work. Personas — genuinely divergent lenses, not paraphrases:
   • Operator — ships the most usable, conversion-mechanical version.
   • Strategist — optimises for positioning, narrative, and differentiation.
   • Skeptic — builds for the distrustful visitor; removes waffle, proves claims.
   • Done when: three complete artifacts exist, none has seen another.
3. Blind-judge. One fresh judge agent (never a contestant) receives the artifacts labeled A/B/C with personas stripped, scores each against the rubric using the judge skill's evidence-ranked manner, and returns per-criterion scores + specific defects + strongest element of each artifact.
   • Done when: scored table + defect list + graft candidates exist.
4. Champion revision — mandatory, the actual alpha. Take the winner, apply the judge's defect list, graft the named strongest elements from the losers. Never blind-ship the raw winner; never concatenate artifacts. One revision round only.
   • Done when: revised champion addresses every judge defect or logs why not.
5. Verify and hand back. Run the artifact (/verify discipline — render it, drive it) and return the tournament record. Code that ships goes through the normal merge-gate path.
   • Done when: champion observed working + record emitted.

Output format

Return to the caller (and drop beside the artifact as tournament-record.md): scores table (per criterion, per contestant), winner + rationale, defects fixed in revision, grafts taken (from whom, what), and the champion's location.

Calibration

• 3 contestants default; 5 max only when the brief names >3 truly distinct audiences/angles.
• Contestants run at medium effort — diversity + judge + revision recover depth (same doctrine as output-tournament). Judge and revision run at high effort.
• Live benchmark: contestant scores 91/93/92 — a spread under ~5 points is normal and does NOT mean the tournament failed; the revision round is the payoff.
• One revision round. A second tournament on the champion is scope creep — hand back instead.

What this skill is NOT

• Not output-tournament — that competes short creative VARIANTS (copy, taglines, directions); this competes complete built artifacts with file output and worktree isolation.
• Not boardroom/ceo-board — those triangulate a decision; nothing is built.
• Not opus-adversary — that reviews ONE existing artifact; here generation itself competes.
• Not a substitute for loops — for iterative refinement of a known direction, loop instead (ranked above tournaments in the source's own testing).

Hard rules

1. Contestants never see each other's work; the judge never sees personas. Contamination observed to collapse the diversity the pattern depends on.
2. The judge is never a contestant (builder≠reviewer — Anthropic engineer doctrine).
3. Synthesis is grafting into ONE coherent artifact — concatenation is the dominant naive failure.
4. Client-facing words inside the artifact still route through nexus-copywriter; the tournament does not bypass the estate content gate.
5. Declare rubric before generation. Post-hoc criteria let the judge rationalise a favourite.

Provenance

• Source: Angela Jiang via Dream Labs live test — https://www.youtube.com/watch?v=PqlnPcae5OQ
• Vault: Wiki/compete-mode-build-tournament-2026-07-15-ingest.md (scores, ranking vs loops).