You know that feeling when you're reviewing a PR and your brain turns to mush after the 50th file? Yeah. Let's fix that.
I've been using AI assistants for code review for about 6 months now, and it's genuinely changed how I work. Not in some futuristic "AI did everything" way — more like "I can actually focus on the stuff that matters."
The Real Problem With Manual Review
Code review is exhausting because you're juggling multiple contexts at once:
- Understanding the business logic
- Spotting actual bugs
- Checking style consistency
- Watching for security issues
- Looking for performance traps
- Maintaining team standards
Your brain can't do all of that simultaneously. You end up missing things or shipping reviews that miss obvious problems.
How I Actually Use AI For Review
I use Claude (paid version, via API) for a two-stage approach:
Stage 1: Automated scan
I dump the PR diff into Claude with a prompt like:
Review this code for:
- Logic errors or edge cases
- Performance issues
- Security vulnerabilities
- Places where error handling is weak
- Code that violates our team patterns
Focus on actual problems, not style.
Takes 30 seconds. Gets me a focused list of real concerns.
Stage 2: Strategic review
I then look at the PR myself, but I'm specifically hunting for:
- Architectural decisions (does this fit the system?)
- Whether the fix actually solves the root problem
- Testing coverage
- Documentation
The AI handles the tedious mechanical stuff. I handle the judgment calls.
Real Example
Last week, a teammate added a database migration. The code looked fine at first glance — proper column types, reasonable defaults. But Claude flagged: "No index on the new foreign key. If you're going to query by this field, add an index."
Caught something I would've missed. We added it before merge.
This happens almost every review.
Tools That Actually Work
- Claude (paid) - Best at understanding context and explaining issues clearly
- GitHub Copilot (paid) - Built into your workflow, integrates with PRs
- ChatGPT (plus) - Works fine, slightly slower
- Local Ollama - Free, but needs a decent machine
I use Claude because the context window is huge (100k tokens) and it rarely makes up problems that don't exist.
The Setup (10 minutes)
- Get an API key from Anthropic, OpenAI, or wherever
- Create a simple script that reads your diff and posts it to the API
- Get the response back formatted nicely
- Read it, make your own decisions, merge or request changes
If you use GitHub, there are already GitHub Actions that do this. I built a custom one that posts to Slack with the findings.
What Still Needs You
AI review misses:
- Whether the feature actually solves the user's problem
- If the architecture is future-proof
- Team context ("oh, we tried this and it caused X")
- Whether the person should've refactored first instead of patching
That's where your judgment lives. AI handles the mechanical stuff. You handle the wisdom.
Numbers That Matter
For me:
- Average review time dropped from 25 minutes to 12 minutes per PR
- Bugs caught increased (AI catches typos and edge cases I miss)
- Time spent on actual judgment increased (better use of attention)
Your mileage varies. If you're reviewing straightforward code, the gains are smaller. If you're reviewing complex systems, this saves your sanity.
Quick Wins Today
- Grab your last problematic PR that had bugs
- Dump it into Claude/ChatGPT with "review this for bugs and issues"
- See what it catches that you missed
- See what it flags that's not real
That difference is your actual win.
Keep Learning
Want to go deeper on AI tools for developers? I write about this stuff weekly in the LearnAI Weekly newsletter — real workflows, real tools, no hype.
Bottom line: Use AI for the tedious parts. Use your brain for the important parts. Review PRs faster, catch more bugs, keep your sanity.
You're welcome.
Top comments (0)