architecture-review

Explain completed or proposed software work so a human can understand its intent, consequential changes, design choices, and supporting evidence.

Restack downstream stacked PRs and apply the current PR's reviewer feedback in their code.

Read from Linear (GraphQL queries and attachment downloads).

Watch a Loom video — pull its transcript and screen frames so the agent can understand what was said and shown.

Investigate feature implementations and generate comprehensive top-down Markdown documentation with Mermaid diagrams.

Produce a compact, syntax-highlighted pseudo-code gist that explains how a specific code path, class, module, feature, or subsystem works.

Interview the user one question at a time, drilling down into the technical details needed to implement a chunk of work.

Produce a functional spec — a document that defines what a system does and how it behaves from the user's perspective, with no source code, types, or…

Break down implementation plans into detailed, self-contained phased plans where each phase is a single focused commit.

Slice a phased or staged plan into a stack of PRs so the human can review and merge them in parallel as later PRs land.

Drive a single stage from a stages document all the way to a stack of reviewed PRs, autonomously.

Drive a multi-phase plan to completion autonomously by spawning fresh-context subagents — one per phase — that land each phase as a stacked PR.