Duck Critic

Run a producer-critic loop, the way the native Rubber Duck works: you keep producing the work and gate it at checkpoints with a second-opinion critic, then revise until the critic passes.

This is not "hand the whole task to a reviewer subagent". You stay the producer and own the plan, the implementation, and the tests. At high-leverage checkpoints you pause and send your current artifact to a read-only critic from a different model family, take in its findings, and continue. Prefer the native GitHub Copilot CLI Rubber Duck when available. Otherwise drive the same loop through the current harness: a VS Code subagent on a different model, a Copilot CLI custom agent, a Claude Code subagent, or a separate top-reasoning model session.

When to Use

• Use when the user asks for rubber duck, ラバーダック, second opinion, 別モデルレビュー, critic review, plan critique, code review, design review, test review, or another-model review.
• Use when the user says "rubber duck で実装して" / "ラバーダックで作って": you implement it yourself and gate your own checkpoints with the critic. Do not delegate the whole implementation to a subagent.
• Use before nontrivial implementation, after drafting tests, before architecture/deployment decisions, or after repeated failures.
• Skip the critic for small, obvious changes — like the native Rubber Duck, consulting is optional and zero rounds is a valid outcome.
• Do not use for an exhaustive audit, or for creating .agent.md files unless the user explicitly asks for a separate agent-file package.

Producer vs Critic Roles

• Producer (you, the main agent): own and keep producing the artifact — plan, code, tests, design. You never hand the whole job to the critic. You decide checkpoints, send packets, reconcile findings, and apply revisions yourself.
• Critic (a different model): a read-only reviewer that only inspects the producer's current artifact and returns severity-classified findings. It never writes files or runs mutating commands.
• This is a gated checkpoint loop, not two agents running at the same wall-clock moment. The producer reaches a checkpoint, hands off to the critic, gets findings, revises, and re-consults — that is where the second model's value comes from. Running multiple critic lanes in parallel at a single checkpoint is fine; the producer and critic taking turns is the loop.