← Articles

// FIELD NOTE

Master PR Reviews with Claude's Address Review Skill

Cory LaNou

Cory LaNou

Master PR Reviews with Claude's Address Review Skill

Master PR Reviews with Claude's Address Review Skill

Overview

PR review comments piling up? One command handles the entire review cycle — fix code, push, watch CI, reply, resolve threads, and request re-reviews. The /address-review skill turns a tedious multi-step process into a single command.

The Problem

You just got review comments on your pull request. Now what? Manually read each one, find the code, fix it, push, reply, resolve the threads, and re-request the review. On a PR with one or two comments, that's manageable. On a PR with 10, 15, 20 comments? That eats hours.

In my last video, I used the /start-issue skill to build a feature from scratch with Claude Code. That PR — PR #22 — already had review comments from Codex. So today I'm going to show you how I handle PR reviews with another one of my favorite skills: the address review skill.

Running the Command

Fire up clauded — that's my alias for claude --dangerously-skip-permissions. I always run in that mode. If you don't have it set up yet, add this to your ~/.zshrc:

alias clauded='claude --debug --dangerously-skip-permissions'

Then type one command:

/address-review 22

That's it. One command, the PR number, and Claude handles the rest. The process is exactly the same whether it's one comment or 50.

How It Works

The address review skill is a full end-to-end workflow for handling PR feedback. It's not just reading the comments and making edits. There's a lot more going on.

Fetch Review Feedback

First, it fetches all the review feedback from the PR. It actually pulls two different types through the GitHub GraphQL API:

  • Review threads — line-specific comments tied to a file and line number. These can be auto-resolved through the API.
  • Pending reviews — general "changes requested" feedback. These can't be auto-resolved. Only the reviewer can approve those.

The distinction matters because the skill handles each type differently at the end of the workflow.

Categorize and Fix

It reads each comment, figures out what's being asked for — code style fix, logic change, doc update — then goes to the file, finds the exact line, and makes the fix.

Here's the key principle: make the minimal change that addresses the comment. Don't introduce unrelated changes. Just fix whatever the reviewer asked for.

Commit and Push

After all the fixes are done, it bundles everything into a single commit and pushes. Clean, atomic, one commit per review round.

Watch CI

Then it watches CI. If CI fails, it pulls down the error, fixes it, pushes again, and watches again. Same feedback loop I showed you in the last video — it doesn't move on until everything is green.

Reply and Resolve

Once CI passes, it replies to each review comment explaining what was fixed. Brief, professional replies — no fluff.

Then it resolves all the review threads through the GraphQL API. This is the part that saves me a ton of time. Manually clicking resolve on every thread in a big PR is tedious. Claude handles it automatically.

Request Re-Review

Here's one of my favorite parts — it knows about bot reviewers. If Codex left a review, it automatically posts an @codex review comment to trigger a re-review. Same with CodeRabbit or Greptile if you use those.

For human reviewers, it requests a re-review through the GitHub API. It also knows to skip CI bots like github-actions[bot], dependabot[bot], and renovate[bot] — no point requesting re-reviews from those.

The Results

One command kicked off this entire workflow:

  1. Fetched the review comments
  2. Fixed the code
  3. Pushed the changes
  4. Watched CI until green
  5. Replied to the comment
  6. Resolved the thread
  7. Triggered a Codex re-review

All automated. On this PR there was only one comment, so it was quick. But on a larger PR with dozens of comments, this skill saves a massive amount of time. The process scales — one comment or fifty, same single command.

Key Takeaways

  1. One command handles the entire PR review cycle — fetch, fix, push, watch CI, reply, resolve, re-review
  2. GraphQL thread resolution saves massive time — no more clicking resolve on every thread manually
  3. Bot-aware re-review triggers — automatically posts @codex, @coderabbitai, or @greptileai to trigger re-reviews
  4. CI watch loop keeps the PR green — same automated fix-and-push loop through review rounds
  5. Minimal changes only — addresses exactly what the reviewer asked for, nothing more

Get Started

The address review skill is part of the go-workflow plugin in my gopher-ai repo. Clone it, fork it, make it your own.

What's your biggest pain point with PR reviews? The back and forth? The manual resolving? Keeping CI green through multiple rounds of feedback? Drop it in the comments — I might have a skill for that already, or I'll build one.

Want more AI development insights?

Subscribe to the newsletter for weekly tips on using AI in professional development.

Subscribe to Newsletter

// KEEP READING

More articles