PR review for humans prototype

Code review, rebalanced

PR review for humans,
against agent‑generated code.

Agents now open pull requests faster than anyone can read them: a thousand plausible lines, tidy commit messages, green checks. The bottleneck is your attention. This spends it where it counts.

The whole trick: swap the domain.

https://github.com/chrisjowen/pr-review-wheelhouse/pull/3 https://pr.chrisjowen.net/chrisjowen/pr-review-wheelhouse/pull/3

Open the live demo → a real 17-file, +1,079 −205 agent PR, with a pre-generated guided review

Free, instant, no AI: a reading plan

Every public PR, no sign-in. Rules about paths and sizes sort the files before you read a line.

Read first

The changes that matter

The biggest source changes, grouped with their neighbours, two or three files at a time.

Skim

Everything else

Tests, docs, config and small edits, grouped by kind, one click from their diff.

Hidden

Generated noise

Lockfiles, build output, snapshots and minified bundles are counted, then folded away.

Then a second pair of eyes, if you want one

Sign in with GitHub to start a guided review. It walks you through the PR in stages, and a reviewer sits beside the code to talk each one through.

  1. Walkthrough. What the PR does and why, the evidence it claims, where the files went.
  2. Quality pass. Readability only: nesting, long functions, duplication, names. Keep or dismiss each note.
  3. Stops. The important changes, with notes pinned above the lines, alternatives, and the reviewer’s pick.
  4. Skim pile. One line and a risk level for everything else.

You decide. Questions and change requests pin to lines; Finish review drafts your verdict, and you copy it as Markdown. Nothing is posted anywhere.

A stop in the guided review: a reviewer note pinned above the lines it discusses
A stop: notes pinned above the lines they discuss.

Why a human-first reviewer

You stay the reviewer

The AI proposes, you dispose. It can’t hide a source file as “generated” or skip the biggest change: its plan is held to the real file list.

Any model

The engine is model-agnostic: cheap models through OpenRouter, OpenAI, or self-hosted. Bring your own key for unlimited reviews.

Small free allowance

Signed-in reviewers get a small daily AI allowance, with a hard cap per person and overall, so nobody gets a surprise bill.

From GitHub in one click

Drag this to your bookmarks bar, then click it on any GitHub pull request: Review for humans