Code review with AI is everywhere right now. But most takes are trash: "Use Claude to review your PRs!" — sure, if you want a tool that misses the gnarly stuff and rubber-stamps mediocrity.
The real value? AI as a filter, not a judge.
The Problem With "AI Code Review"
Throwing your PR at an LLM and hoping for insights is like asking your rubber duck to spot security bugs. It's good at spotting:
- Obvious naming issues
- Missing error handling in the happy path
- Async/await footguns
- Dead code
It's terrible at spotting:
- Whether this belongs in this service (architecture)
- Performance implications you can't see in isolation
- Why the original author wrote it this weird way (context you're missing)
- What your actual users will complain about
What Actually Works
Use AI as your pre-review pass. Before a human touches your PR:
-
Lint the code (by hand? no, use Copilot/Claude for this):
- Run it through an AI with this prompt: "Flag any obvious issues: naming, error handling, TODOs, security concerns. Don't judge architecture — just the surface stuff."
- Takes 30 seconds, catches the "oh I forgot to close this file descriptor" bugs.
-
Generate a summary:
- Feed the diff to Claude: "What does this PR do? List the key changes and any obvious gotchas."
- Now the human reviewer doesn't start cold. They know what they're looking for.
-
Ask about specific concerns:
- You know this code is gnarly? Ask: "What are the failure modes if X goes wrong?"
- AI can brainstorm edge cases faster than you can think them.
Real Example: The Database Migration
I was pushing a migration that dropped a column after a deprecation period. Basic stuff, right?
Threw it at Claude:
"What could go wrong if code that reads this column hasn't fully rolled back yet?"
Claude: "If service A hasn't deployed the removal code, it'll error when the column vanishes. Reads return NULL → crashes. You should add a default or handle NULL."
Would I have caught that alone? Eventually. Would my human reviewer? Maybe. But now we all see it immediately.
Tools Worth Using
GitHub Copilot + PR comments:
- Hit Copilot in your editor, it suggests review comments right in the diff
- Fast, surface-level, prevents "oh we don't use console.log in prod" conversations
Claude or GPT-4 in a separate window:
- Paste the diff, ask specific questions
- Better for "what's the reliability story here?" stuff
Phind or Tabnine:
- Built specifically for code, faster on smaller diffs
- Good for quick checks before you tag humans
Don't pay for a dedicated "AI code review tool" — they're just wrappers around these APIs charging you 10x.
The Hard Truth
AI code review fails when:
- You use it to avoid thinking
- You trust it on architecture decisions
- You skip actual human review (especially for risky changes)
- You don't understand why it flagged something
It wins when:
- It handles the boring pass first
- Humans focus on the interesting problems
- You stay skeptical and verify its suggestions
- You're asking it specific questions, not "review this"
One More Thing
If you're building this as a personal workflow, stick it in a simple script. Shell out to the API, pipe your diff in, get results in Slack. Takes an afternoon.
Learn more about AI workflows and hands-on tools in the LearnAI Weekly newsletter — practical guides on this stuff, actually useful.
Stop letting AI replace your judgment. Use it to speed up the parts that bore you.
Top comments (0)