Review your own PR like a stranger would.
Feed Claude the diff plus the PR description. It pushes back on unclear naming, missing error paths, and the part where you said "temporary" three months ago.
Bootstrap a repo without the boilerplate.
Drop Claude into an empty directory and it'll ask six questions, then scaffold the stack, config, and README you'd have typed by hand. Skips the parts you always delete anyway.
Turn an API reference into a typed client.
Point Claude at the docs URL. It generates typed request/response interfaces, a thin client wrapper, and the three edge cases their docs forgot to mention.
Track a bug back to the first bad commit.
Paste a failing test and the symptom. Claude runs `git bisect`, inspects diffs, and narrows the culprit. Works best on repos where history is clean.
Generate tests that actually fail when the code breaks.
Instead of mirroring the implementation, Claude infers the public contract, picks edge cases, and writes tests against the interface — the kind that survive a refactor.
Refactor in three commits, not one.
Claude splits a hairy refactor into a preparatory pass, the move, and a cleanup — each passing tests independently. Bisectable if something breaks later.