Agent-First Code Review Workflow

AI Engineer's principle

The workflow for verification and code review is agent-first: Claude does the first-pass review, the agent is given a way to try the experience like a real user, and the human is responsible for checking the agent's verification evidence.

Run an agent-first code review loop where Claude reviews changes first, the agent verifies the experience like a real user, and you check its evidence before the PR is corrected.

The source describes a workflow the creator's team runs; the cadence of running it and the specific shared files to write are the user's to choose. All action text is source-backed; the two parameters only fill in the schedule and the file names.

1. 01

   Have Claude do a first-pass review of code changes.

   Recurring · on your cadenceDone when: Claude has completed a first-pass review of the code change.

2. 02

   Give the reviewing agent a way to actually try the experience out similar to a real user: end-to-end tasks, opening up the experience, checking it in a browser, running tools, and doing linting.

   Recurring · on your cadenceDone when: The reviewing agent has a working way to try the experience out like a real user.

3. 03

   Set up hooks that run on any edits, study the diffs, and drive a loop that corrects the PR as quickly as possible when a bug or gap in functionality is discovered.

   Setup · onceDone when: Hooks are in place that run on any edits, study the diffs, and drive a correction loop for the PR.

4. 04

   For UI changes destined for the browser, have the agent test the change in the browser and generate a video of the experience where it clicks around, to build confidence that it works.

   Recurring · on your cadenceDone when: The agent has tested the UI change in the browser and generated a video of the experience where it clicks around.

5. 05

   For mobile or desktop experiences, have the agent generate screenshots or videos of that experience.

   Recurring · on your cadenceDone when: The agent has generated screenshots or videos of the mobile or desktop experience.

6. 06

   Check the agent's verification evidence yourself — for UI work used by a user in the browser, have the agent generate videos or screenshots of the browser experience to build confidence it works.

   Recurring · on your cadenceDone when: You have checked the agent's verification evidence yourself.

7. 07

   Codify your team's best practice and taste in shared files such as [which shared files will you codify your team's best practice and taste in] where the agent can consult them during code review.

   Setup · onceDone when: Your team's best practice and taste are written down in the shared files the agent can consult during code review.

8. 08

   Review whether the agent-first review loop is still working for your team: is Claude's first-pass review catching issues, is the agent's verification evidence giving you confidence, and are the shared files still reflecting your team's best practice and taste?

   Review · on your cadenceDone when: You have reviewed whether the loop is still working and noted any changes to make.NuthinButta scaffolding