lobster-coordinator
Production-grade 3-layer agent orchestration with dual-blind verification and automated model routing.
skill install https://www.promptspace.in/skills/lobster-coordinatorWhat it does
The Lobster Coordinator is a high-fidelity orchestration framework reverse-engineered from production-grade agentic platforms like Claude Code. It moves beyond simple prompting by implementing a 3-layer architecture: AgentTool for task delegation, Swarm for team messaging, and Coordinator for global task decomposition and aggregation.
Why use this skill
Building reliable multi-agent systems is difficult. This skill provides a battle-tested blueprint that eliminates "agent loop" failures through secondary verification layers. It prevents hallucinations and technical debt by enforcing strict multi-stage reviews extracted from the Santa Method and Superpowers frameworks.
- Two-Stage Review: Automated checks for Spec Compliance followed by Code Quality.
- Santa Method: Dual-blind verification where two independent reviewers must approve code before merging.
- Model Routing: Optimize costs by routing tasks to model tiers based on complexity (Cheap/Standard/Elite).
- Status Handling: Robust state management for agents that are DONE, BLOCKED, or NEED_CONTEXT.
Supported tools
This skill is designed for advanced AI environments like OpenClaw, Claude Code, or Cursor, specifically utilizing capabilities like sessions_spawn for isolated sub-agent execution.
Use cases
- Decompose complex software tasks into parallel sub-agent workflows.
- Enforce dual-blind 'Santa Method' code reviews for production releases.
- Optimize API costs by routing sub-tasks to smaller, cheaper models.
- Isolate tool permissions across different specialized agent roles.
Example
Prompt
Output
[Coordinator] Decomposing task into 3 sub-units... [Worker-1] Status: DONE | [Review-A] Spec: PASS | [Review-B] Quality: PASS [Santa-Verify] Independent Reviewer C: PASS | Independent Reviewer D: PASS [Coordinator] All checks passed. Merging 142 lines of verified code to main.