What problem does it solve? Large refactors that span multiple architectural layers are too big to design and execute in a single session, leading to inconsistent code, missed library API changes, and abandoned half-finished migrations. This Skill freezes a validated design plan first, then dispatches the mechanical work in small atomic chunks to subagents. ## Core Features & Use Cases - Validated plan generation: Produces a structured design plan (sections 0-13) with mandatory context7 validation of every third-party library claim and a harmonization pass that enforces one canonical template across the layer. - Copy-paste orchestrator prompt: Decomposes the plan into a chunk queue (one chunk = one atomic commit) and generates a ready-to-paste orchestrator prompt that dispatches Haiku executor agents, reviews each chunk, and escalates contradictions to a Sonnet investigator agent and the user. - Self-improving retrospective loop: A final analyzer agent writes a retrospective report, appends observations to a learnings ledger, and proposes template or agent improvements via a rule-of-three promotion mechanism. - Use Case: You need to redesign your FastAPI router layer and push business logic down into an application layer across 40 files. Run this Skill to produce the frozen plan and orchestrator prompt, then paste the orchestrator into a fresh session to execute the refactor chunk by chunk with atomic commits. ## Quick Start Ask the AI to orchestrate a top-down refactor of a specific layer, for example: "Orchestrate a layer-by-layer refactor of the routers layer and generate the plan and orchestrator prompt."
