junior-to-senior

junior-to-senior is an adversarial review skill that treats any agent-generated plan, design, or architecture as "junior" work and constructs a rigorous "senior" reviewer to tear it down and rebuild it. It exists to catch the two ways AI-written plans fail: "fog" (high-level hand-waving with no interfaces, data shapes, failure handling, or named libraries) and "tunnel" (diving into function signatures while missing product vision, scope, and rationale). The core discipline is evidentiary: every finding must cite a file and line for codebase claims, or a dated primary source for best-practice claims, and any claim that cannot be verified via web research is explicitly labeled as unverified training data.

The workflow runs in phases. Phase 0 freezes the artifact under review so critique targets a fixed text. Phase 1 builds the senior reviewer by extracting 2-5 load-bearing technical domains, doing live code research (conventions, lockfile versions, prior art, real constraints) and web research for deprecations, supersessions, and hard-won lessons, ideally in a context-isolated subagent so the review is adversarial rather than self-confirming. Phase 2 diagnoses altitude (fog, tunnel, or mixed), and Phase 3 delivers findings graded as blocker, major, or minor, each with evidence and a concrete fix. Supporting reference files provide the review rubric, a vague-word blacklist, and a research playbook.

The skill targets engineers and teams using AI agents to produce technical plans who want a staff-level sanity check before committing. Typical use is invoked by phrases like "senior review" or "review this like a staff engineer," or automatically before acting on any agent-written plan. It is process guidance rather than executable tooling, and its quality depends on available codebase and web access.