This is a PR review assistant built specifically for PyTorch contributors. It works three ways: point it at a GitHub PR number, run it on your local branch, or invoke it directly from GitHub Actions. The philosophy here is refreshingly blunt: only report problems, no praise, and if you're uncertain whether something applies, spawn a sub-agent to go read the actual code instead of guessing. It checks what CI can't catch: backward compatibility implications, thread safety, test coverage adequacy, whether you're using the right dispatch keys. The checklist is extensive and PyTorch-specific, things like device guards and TensorIterator usage patterns. After drafting findings, it fact-checks itself by sending sub-agents to verify each claim independently.
npx -y skills add pytorch/pytorch --skill pr-review --agent claude-codeInstalls into .claude/skills of the current project.
Select a file.
github/awesome-copilot
microsoft/win-dev-skills
github/awesome-copilot