Marshal is a session-scoped orchestrator that turns a user direction (audit, fix, map, research) into a plan of actions and executes them by chaining skills and tools. It focuses on bounded, single-session problems: investigate, plan, execute, verify, and report.
Use Marshal for multi-step but bounded tasks that need investigation and cross-skill coordination (e.g., 'audit the UX', 'fix the auth bug', 'map the API surface'). Do NOT use for single-file edits, long-running multi-session campaigns, or parallel work better suited to a fleet controller.
Designed for agents that can load local skills and run tool calls (Claude Code-style harnesses). Works with orchestration-capable agent runtimes that support spawning sub-agents and enforcing timeouts.
Marshal is a single-session meta-orchestrator skill that takes a direction, chains existing skills, and executes multi-step tasks end-to-end. No bundled scripts — purely instructional SKILL.md. Well-structured with clear phases (Understand, Plan Chain, Execute, Report, Learn), quality gates, timeout handling, and fringe case guidance. The skill is text-only with no executable components.
Clean skill with no security concerns. Pure instructional text — no scripts, no network calls, no file operations. Well-documented with clear output contracts (Marshal Report format, HANDOFF block). Could benefit from separating into references/ for longer examples, but the single-file approach works for this scope.
