idea-duel

name	idea-duel
description	Lean dueling-wizards ideation tournament — the two model stacks independently study a project, generate and winnow ideas, blind cross-score each other's, and probe for shared blind spots; the orchestrator synthesizes a consensus matrix whose winners become draft work items. Use when the user wants the strongest next ideas for a project ("what should we build next", "run an idea duel", "most valuable improvements"). Four dispatches, discretionary quota spend — an upstream generator feeding the capture pipeline, typically offered from /discussion; never part of /do.
argument-hint	[project or area to ideate on, plus any focus or constraints]

Idea Duel — lean dueling wizards

Single-model brainstorming has one flaw that matters: the model that generates an idea also evaluates it. The duel splits those jobs across the two stacks and makes convergence the quality signal — where two differently-biased models agree blind, the idea is probably good; where they trash each other's, it's suspect. Four dispatches, ~100k tokens at the pinned efforts — about one review pass.

Wizards (pinned)

• Claude wizard: sub-agent, model: opus, thinking/high effort.
• Codex wizard: dispatched directly via Bash (timeout 600000 ms) — codex exec -m gpt-5.6-sol -c model_reasoning_effort="low" --sandbox read-only --ephemeral --skip-git-repo-check -C <repo root> -o <scratchpad>/codex-idea-wizard-<n>.md "<prompt>". Each dispatch is a fresh session that knows nothing of this conversation — the prompt carries the brief and names both files below by absolute path; Codex reads them itself.
• Both follow references/instructions.md (stack-neutral) and write the formats in references/idea-wizard-output.md — both under this skill's base directory.
• If a participant degrades (fallback model, truncated output), retire it and report a single-wizard result honestly — never contaminate scoring.

Phases

0. Frame (orchestrator, no dispatch): one-paragraph brief — the project, the focus, and the ACTUAL constraints (scale, maturity, appetite); artifact dir ./tmp/idea-duel/<slug>/. An honest brief is most of the value: omit the real constraints and you buy ideas tuned to someone else's context.
1. Generate (2 dispatches, parallel, blind): each wizard independently studies the project → generates ~10 candidate ideas → self-winnows to a top 3, self-ranked with reasoning → writes ideas-<stack>.md. Neither sees the other.
2. Cross-score + blind-spot probe (2 dispatches, parallel): each wizard gets the brief and BOTH idea files, and — before seeing any scores — scores the opponent's 3 (0–1000, one-line verdict each; the number is a commitment device, not a measurement) and nominates blind spots: the strongest idea NEITHER list contains. Writes scores-<stack>.md.
3. Synthesis (orchestrator, no dispatch): assemble the matrix — idea × self-rank × opponent score × verdict — and classify:
   • consensus (opponent score ≥ 800, or both stacks listed the same theme): winners;
   • contested (score delta from self-rank expectation ≥ 250): resolve by staging ("yes, but v1 first") or, when genuinely zone 0–1, flag for the dialectic skill rather than settling it here;
   • killed: neither side defends it — drop with one line of why. Check the blind-spot nominations for convergence — two stacks independently nominating the same missing idea is the strongest signal the duel produces; promote it to the winners regardless of the matrix. Emit DUEL_REPORT.md: the matrix, the reasoning, and a dependency-ordered sequence (enablers first). Winners become status: draft work items under ./tmp/<idea-slug>/item.md, each carrying its duel evidence — both scores, the verdict line, and what it beat — pre-seeded into the item's Justification and refs/ (link DUEL_REPORT.md). The Socratic gate then serves as the tournament's third adversarial filter, and calibrates down accordingly: an item with documented duel evidence earns a fast pass unless its premise has a hole the duel never tested.

Rules

• Wizards are read-only toward the repo: cite, never edit. No dispatch sees the opponent's scores before its own are written.
• The orchestrator never adds its own ideas to the matrix — it frames, synthesizes, and judges; generation belongs to the wizards.
• Every idea in the report carries its origin, both numbers, and the verdict — survivors must be traceable back through the duel.