← ClaudeAtlas

swarmlisted

Launching multi-agent parallel work with the Agentic SDLC. Use when a task benefits from decomposition into parallel subtasks.
rmzi/portable-dev-system · ★ 24 · AI & Automation · score 53
Install: claude install-skill rmzi/portable-dev-system
# /swarm — Agentic SDLC Six-phase workflow for decomposing, dispatching, and validating parallel work across agent teams. Each phase shows the concrete tool calls needed to execute it. ## Delegation **If you are not the orchestrator**, spawn one to execute this workflow. Agents terminate when they return output, so **the parent must handle the human approval gate** — not the orchestrator. **Two-phase delegation pattern:** ``` # Phase 1: Orchestrator runs grill, produces plan, returns it plan = Agent(subagent_type="pds:orchestrator", name="orchestrator-plan", prompt="Run /pds:grill for: <task description>. Tier: <tier>. Produce a plan with acceptance criteria. Return the plan — do NOT proceed to decomposition. Write swarm state files (.claude/swarm/phase, .claude/swarm/tier) before returning.") # Parent relays plan to human, gets approval/override, then: # Phase 2+: New orchestrator executes the approved plan through all remaining phases Agent(subagent_type="pds:orchestrator", name="orchestrator", prompt="Plan is approved by human. Execute /pds:swarm Phases 2-6 for: <context>. <paste approved plan + acceptance criteria here>. Proceed through all phases without stopping for approval.") ``` If no tier is specified, the Phase 1 orchestrator MUST run `/pds:grill` first to determine the tier. Grill is mandatory before any swarm — it validates requirements AND recommends a tier. The orchestrator ha