add plan-prd skill

[mikigraf]

/plan-prd: Decompose reviewed plans into parallel, agent-ready PRDs

New skill in the plan-* family that sits at the end of the review chain. Takes a plan that's already been through CEO, design, and eng review and decomposes it into a set of self-contained PRD files, one per independent work unit, designed for concurrent development by multiple agents.

Where it fits in the workflow:
/office-hours > /plan-ceo-review > /plan-design-review > /plan-eng-review
|
/plan-prd > N independent PRDs + orchestration plan
|
Parallel implementation (one agent per PRD per lane)

What it adds beyond the existing review chain:

• Decomposition into separate files. The reviews produce one plan. This skill produces N PRD files, each scoped to one work unit with its own user stories, acceptance criteria, tech spec, test plan, and implementation checklist.
• Shared interface contracts. Typed contracts between PRDs so parallel agents code against interfaces, not each other's incomplete code. Each PRD defines what it exposes and what it consumes.
• Coverage matrix. 4 mandatory cross-reference iterations over all source documents (coverage audit, interface consistency, completeness sweep, dependency validation) to prove every requirement landed in a PRD or was explicitly excluded.
• Orchestration plan. Wave-based execution order, merge strategy, stub strategy for maximum concurrency, risk register.

What it intentionally does not duplicate:

• Parallelization lane analysis (leverages eng review's existing worktree strategy as starting point)
• Contradiction resolution between reviews (the sequential review chain already handles this)
• Scope, design, or architecture decisions (those are the upstream reviews' jobs)

Output: PRD files and an orchestration plan written to ~/.gstack/projects/{slug}/prds/. 867 lines, ~9.8k tokens.