What problem does it solve?
Complex multi-step tasks often fail when executed in a single pass because later steps depend on earlier results and errors compound silently. This Skill orchestrates a task as ordered, dependent steps run by isolated sub-agents, passing context forward and verifying each step with an independent LLM-as-a-judge before proceeding.
Core Features & Use Cases
- Sequential Orchestration: Decomposes a task into ordered subtasks with a dependency graph, then dispatches implementation sub-agents step by step with context summaries from prior steps.
- Meta-Judge Verification: Runs a meta-judge in parallel with each implementation agent to generate a tailored evaluation specification, then an independent judge scores the step with up to 3 retries.
- Per-Step Model Selection: Assigns haiku, sonnet, or opus tiers per step based on scope, complexity, and risk, with escalation rules and cross-provider model mapping.
- Use Case: Refactoring a shared interface across implementations and consumers — the interface change is verified before implementations update, and implementations are verified before callers are touched.
Quick Start
Ask the agent to run do-in-steps with a task description such as "Refactor UserService class and update all consumers", optionally adding --strict or a --model override.
