/diff-walkthrough

Present a changeset reorganized for reviewer comprehension — not in file-tree order, not a raw diff dump. This skill explains what happened; it does not review. Findings belong to /review or /basic-review.

Gather the diff

Accept a branch, a range, staged/uncommitted changes, or a PR reference (with gh pr diff when the remote is reachable). If the user did not say which changeset they mean, stop and ask. Never guess from recent history or fall back to a silent default — a walkthrough of the wrong diff is worse than no walkthrough.

Group changes for comprehension

Do not present files in alphabetical or tree order. Reorganize into sections ordered by reviewer value:

1. Core logic — new behavior, algorithm changes, state transitions, API surface changes. Show the real hunks with surrounding context.
2. Wiring & integration — route registration, dependency injection, config plumbing that connects the core logic. Condensed: enough to confirm correctness.
3. Boilerplate & mechanical — renames, import reordering, generated code, formatting, type re-exports. Summarize as a list of files and stats; no inline hunks unless something is genuinely surprising there.

Lead with core logic: attention is freshest at the top.

Distill dense logic into pseudocode

When a core hunk is dense — nested conditions, state machines, retry/backoff, multi-step transforms — add a few lines of pseudocode next to it that strip syntax and error handling to expose the essential control flow. Only for hunks that are genuinely hard to scan; straightforward changes need no mirror.

Trace surprising behavior on a concrete example

When a hunk changes behavior in a way that is hard to predict from reading it — reordered operations, new short-circuits, altered edge cases — pick one small, realistic input and walk it through the old and new code paths side by side, highlighting the step where they diverge and the observable outcome. Reserve this for genuinely surprising changes.

Call attention to tricky things — sparingly

When a hunk hides something risky or easy to miss, mark it with a short tag (Subtle, Breaking, Race condition, Perf) and one sentence. Overuse destroys the signal; a few tags per walkthrough is the ceiling.

Tone

Write reviewer-facing commentary, not a changelog: why it changed, how files interact ("the new validator in core.py is invoked by the route added in routes.py"), anything the diff alone does not make obvious. One or two sentences per note. No findings, no verdicts, no "LGTM".

Adapted from cursor/plugins pr-review-canvas/skills/pr-review-canvas/SKILL.md (MIT, © Cursor) — the canvas-SDK presentation layer replaced with markdown conventions.