swarm

What problem does it solve? Running parallel AI workers without fixed judging rules leads to biased picks, overlapping work, and reports that hide gaps. This Skill enforces a disciplined fan-out workflow where the done predicate, judging rule, and disjoint scopes are all written down before the first worker spawns. ## Core Features & Use Cases - Four swarm shapes: coverage splits a scope into slices, race runs one brief N times, gauntlet runs N distinct checks against one artifact, and arena has N candidates build the same thing for rubric-judged comparison. - Pre-committed judging: a frame file fixes the done predicate, worker slices, and rubric before results exist, and a separate read-only judge delegate scores arena candidates criterion by criterion. - Disjoint execution: each worker gets its own worktree or output file, pinned SHAs and measurement methods, and returns a one-line verdict while full detail goes to its own file. - Use Case: You want three independent implementations of a parser compared fairly. Invoke the arena shape, let three candidates build in separate worktrees, have an Opus judge score them against a fixed rubric, then graft the best parts into one design. ## Quick Start Invoke the swarm skill with a shape, a done predicate, and a worker count, for example by typing /exo:swarm arena "a retry wrapper for the HTTP client" 3.