What problem does it solve?
Large tasks that exceed a single model's context window or require parallel research across many subtasks often fail due to context leaks, silent partial failures, or judgment applied too late. This Skill enforces a disciplined three-tier loop that splits work into stateless briefs, verifies every result, and consults a stronger advisor model only at commitment boundaries.
Core Features & Use Cases
- Three-tier model team: Orchestrator plans and verifies, cheap parallel workers execute subtasks, and a stronger advisor reviews the plan and final deliverable.
- Stateless worker briefs: Every dispatch carries full inputs and acceptance criteria inline, written to temp files to prevent shell injection and context leakage.
- Mandatory verification: Each result is judged PASS, FIX (redispatched with named failure), or ESCALATE against its own acceptance criteria.
- Advisor as critic: Two mandatory consults (plan review and taste pass) plus mid-loop escalation, with every note applied or explicitly rebutted.
- Use Case: Research twelve competitors in parallel, synthesize findings into a report, and have a stronger model review both the research plan and the final draft before delivery.
Quick Start
Ask your coding agent to orchestrate this task across a model team using the advisor-worker loop.
