Just realized something—our team’s code review checklist has been sitting in a Confluence doc, mostly ignored 😅 But last week I turned it into a structured AI prompt, and it’s already catching things we used to miss.
Here’s how I did it in under 30 minutes:
- **Pull the checklist items** – ours had things like “error handling present,” “no hardcoded configs,” “logging consistent.”
- **Add context** – I prefaced with our team’s language (Python/JS), any style guide links, and key file paths.
- **Set the tone** – asked the assistant to “act as a senior reviewer” and flag issues with severity levels.
- **Test it** – pasted a recent PR diff and refined until it gave actionable feedback.
Now we drop the prompt into any chat when reviewing code. It’s not a full replacement, but it’s a killer first pass.
Anyone else tried this? Would love to swap examples, especially for security or performance checklists.
— Jason
Let's build better workflows.
Interesting approach. I've been doing something similar for performance-related reviews but added a quantification layer. After converting our checklist, I ran it against a corpus of past PRs where we later discovered performance regressions.
For example, one prompt item evolved from "check N+1 queries" to "analyze the diff for new loops over database result sets, especially in nested structures, and flag any that lack batch loading patterns." The specificity matters because generic warnings get ignored.
The caveat is you need to validate the AI's feedback against real data initially. I logged its suggestions from 50 old reviews and cross-referenced with actual production metrics - about 30% were false positives focusing on micro-optimizations. Trimmed those items and the prompt became much more credible with the team.
Would be curious to see how you handle severity calibration. Do you weight security items higher automatically, or does the prompt try to assess context?