The Linus Method: How we simiplifed RFC reviews (www.devashish.me)

A team discovered a simple but powerful trick for getting honest, useful architecture reviews from large language models: ask the model to “review this as Linus Torvalds would.” After routine AI reviews produced polite approvals that missed systemic flaws, invoking the Linus persona produced blunt, consequence-focused feedback (e.g., flagging distributed-monolith coupling via shared DBs and warning about deployment/backlog pain). They standardized the approach with a short prompt—“Review this [RFC/Architecture/PRD] as Linus Torvalds would, focusing on long-term maintainability. Context: [domain, constraints, goals]. No profanity.”—and integrated it into their RFC workflow. Technically, the method works because major LLMs have absorbed thousands of real-world Torvalds-style reviews and rants; the persona activates that coherent review philosophy (long-term thinking, eliminating special cases, structural elegance) and sidesteps sycophantic, surface-level praise. It’s most effective for full-context architecture reviews, design docs, and feasibility assessments and less suited to early brainstorming, first drafts, or delicate junior proposals where bluntness can demoralize. The practical implication: teams catch architectural debt earlier, raise technical standards, and create a portable, model-agnostic review cadence—provided reviewers supply complete context and temper the approach for sensitive situations.