Code reviews backed by a dozen engineering classics

What it does

brooks-lint is a skill pack for Claude Code, Gemini CLI, and Codex CLI that prompts LLMs to audit code against six production and six test-suite “decay risks.” Each finding must include a symptom, a consequence, a remedy, and a citation to one of twelve canonical books—think Refactoring, The Mythical Man-Month, or Working Effectively with Legacy Code. The output is a Markdown report with a 0–100 health score, severity emojis, and, in architecture mode, a Mermaid dependency graph color-coded by risk.

The interesting bit

The project treats classic software engineering literature as the system prompt’s guardrails. Rather than asking an LLM to “review this,” it constrains the model to a specific diagnostic ontology—cognitive overload, change propagation, domain model distortion, and others—derived from the books. The README claims this produces far more consistent structured output than raw Claude, though the benchmark is self-administered.

Key highlights

• Bibliographic citations: Every finding traces back to a specific classic book and concept (e.g., Fowler’s Refactoring on Divergent Change).
• Six analysis modes: PR review, architecture audit, tech-debt assessment, test health, a combined dashboard, and a full-sweep auto-fix mode.
• Visual architecture audits: Generates Mermaid dependency graphs with nodes colored red, yellow, or green based on finding severity.
• Cross-language: The README claims it works with any language, since it operates via LLM analysis rather than static parsing.
• Plugin packaging: Ships as native skills and plugins for Claude Code, Gemini CLI, and Codex CLI.

Caveats

• Not standalone: It is a prompt-and-skill framework, not an independent tool; you must already be running Claude Code, Gemini CLI, or Codex CLI.
• Self-reported benchmarks: The 94 % pass-rate claim comes from the project’s own comparison against raw Claude, not an independent study.
• LLM-dependent: Output quality and consistency rely entirely on the underlying model’s willingness to stay inside the guardrails.

Verdict

Teams already using Claude Code or similar LLM CLI tools and who want more structured, book-traceable architectural feedback should try it. If you need deterministic static analysis or don’t use one of the supported AI assistants, this adds nothing.