Claude Orchestrator Skill — AI skill for Claude Code

Multi-Model Orchestrator Skills

Two native skills that make a frontier model the CEO of a cross-vendor agent team. Same doctrine in both; the seats are chosen by whichever model is driving.

Two invariants hold across every row:

• Whoever built it never certifies it, and the reviewer always comes from a different family than the builder. This is the whole point. Two runs of the same model make correlated mistakes, because the same training produces the same blind spots, so a second pass mostly agrees with the first. A model from a different lab fails in different places. The corollary lives in both skills: agreement between the two is weaker evidence than it feels, and disagreement is a signal to slow down rather than to average.
• The driver makes every final call. Advisers and reviewers produce evidence and disagreement. They never hold the decision, and a review is never a rubber stamp or a veto.

Both share one North Star with two goals in order: maximize the driver's performance on multi-disciplinary work, and minimize the tokens spent getting it. Quality wins when the two genuinely collide, but the bet is that they rarely do. To that end the skill grants the driver two things it does not otherwise use freely: a frontier peer from another model family to think alongside it and to check its work, and a roster of cheaper but still capable workers to hand bounded work to whenever that saves tokens without costing quality. A short precedence l