Teaching Claude Code to Review Like a Senior Dev

What it does

It’s a curated set of review guidelines—over 16,000 lines across 20+ languages and frameworks—that plugs into Claude Code as a “skill.” The core prompt is only ~190 lines; language-specific guides (200–1,000 lines each) load on demand when you ask it to review React, Rust, Python, or whatever else is in the PR. Every finding gets a severity label from blocking down to praise, and the review follows a four-phase workflow from context gathering to final summary.

The interesting bit

The clever part is the progressive disclosure: instead of dumping a massive system prompt into every session, it keeps the base skill lightweight and pulls in expertise only when needed. It also explicitly separates what a human reviewer should catch from what linters already handle, which is a rare bit of meta-awareness for an AI tool.

Key highlights

• Covers modern stacks: React 19/Next.js, Vue 3.5 Composition API, Svelte 5, Rust, Go, Zig, Swift, Kotlin, and roughly a dozen more, plus cross-cutting architecture and performance guides.
• Severity taxonomy: blocking, important, nit, suggestion, learning, praise—so feedback has actual weight.
• Security-first angle with dedicated checklists per language ecosystem.
• Includes a PR complexity analyzer script (pr-analyzer.py) and review templates in assets/.
• Explicitly collaborative tone: questions over commands, suggestions over mandates.

Verdict

Worth a look if you use Claude Code and want consistent, structured reviews across a polyglot codebase. If you don’t use Claude Code, this is just a very large Markdown repository of review checklists you could theoretically read yourself—but that misses the point.