What you get.
A focused Claude Skill aimed at one thing: understanding every unsafe in your Rust code well enough to defend it in PR review. No vibes, no "looks fine to me." It reads the code, the surrounding lifetime, the docs nearby, and either signs off or rewrites it.
Used by 14 crate authors at the time of writing — including two on tokio's auxiliary crates and one on a wgpu fork.
- Walks every
unsafeblock, classifies it (FFI, raw-pointer arithmetic, lifetime extension, repr-cast). - Cross-references with rustdoc and pulls the SAFETY comment if one exists, flags missing ones.
- Suggests safe rewrites where possible — usually
NonNull,Pin, or a typed wrapper. - Outputs a Markdown audit table you can paste into PRs.
Drop the .md in .claude/skills/, restart Claude, and prompt it on any crate path.
Spec.
Scanned & signed.
Reviews (3).
Caught a transmute<&[u8], &[u32]> in our parser that nobody had touched in 3 years and that was technically UB on big-endian. The audit table format is exactly what I'd write myself in PR review, just faster. Cheap at twice the price.
Skill's prompt is well-tuned — it doesn't try to "fix" unsafe that genuinely needs to stay (FFI). It explains why and moves on. Big improvement over generic "is this safe?" prompts.
Good skill. Would be 5★ if it included a Cargo.toml manifest scan for unmaintained deps. Maybe v2.5? Either way, paid for itself on the first run.