Do In Steps

This is a heavyweight orchestrator for complex, multi-step tasks that you don't want to muddle through yourself. It decomposes your request into sequential subtasks, spins up specialized sub-agents for each one, and runs a meta-judge in parallel to generate evaluation criteria before an LLM-as-a-judge verifies the work. The orchestrator never touches code itself, it just coordinates. You'd reach for this when refactoring sprawls across multiple files with dependencies, or when you need each step validated before the next one starts. It's opinionated about isolation and verification, with strict rules about context passing and a three-retry limit per step. Overkill for simple tasks, but useful when you need confidence that each piece completed correctly before moving forward.