DEV Community

Patrick
Patrick

Posted on

Using Claude Code for PR Review: What 30 Days of Real Usage Taught Us

#claudeai #productivity #devtools #git

Title: Using Claude Code for PR Review: What 30 Days of Real Usage Taught Us

Tags: claudeai, productivity, devtools, git

Pull requests are where bugs sneak through. Most teams rely on tired reviewers skimming 800-line diffs at the end of a sprint.

We've spent the last 30 days using Claude Code as a first-pass reviewer before human review. Here's what actually works.

The Basic Workflow (Week 1)

Before submitting a PR, run this prompt in Claude Code:

Review this diff as a senior engineer. Focus on:
1. Logic errors and edge cases I might have missed
2. Performance implications
3. Places where the intent isn't clear from the code alone

Diff:
[paste diff]
Enter fullscreen mode Exit fullscreen mode

What surprised us: Claude catches the things humans rush past. Not the clever stuff — the boring stuff. Missing null checks. Off-by-one errors. Functions doing more than one thing.

The Team Workflow (Week 2-3)

Once one person got value, we standardized it. We added three questions to every PR description template:

  1. "Did you run Claude Code on this diff?" (yes/no checkbox)
  2. "What did it flag?" (2-3 sentences)
  3. "What did you disagree with and why?"

That last question is the important one. It turns Claude into a forcing function for thinking about your own code — not just accepting the feedback, but reasoning about it.

What It Doesn't Replace

Claude Code won't catch:

  • Whether the feature is the right feature
  • Architectural decisions that require knowing the full system
  • Context about product decisions that live in Slack, not code

Human reviewers still own those. But human reviewers spend way less time on the tactical stuff now.

The Utilization Reality

Most teams that "rolled out" Claude Code did it the same way: one Slack message, one YouTube link, no structure. 30 days later, 70% of the team hasn't opened it this week.

The PR review workflow is a forcing function. It gives devs a reason to open Claude Code on every single PR, twice a day, in a structured way. Utilization goes from "when I remember" to "this is part of the process."

In our data across teams we've trained: average utilization at 30 days without a forcing-function workflow is 22%. With one (like PR review), it's 61%.

The tool isn't the problem. The workflow is the problem.

Get the full Claude Code Team Playbook (10 modules, free): askpatrick.co/playbook-sample.html

If your team's utilization is stuck, we also run half-day live workshops. More at askpatrick.co.

Top comments (0)