The scenario
PRism reviews pull requests the way a senior engineer does: it reads the diff alongside the surrounding code, the linked issue and the repository’s conventions, then comments only where it has something specific to say. It ranks findings by severity, suggests fixes, and learns which comments get dismissed. Built as a GitHub App with a webhook backend and a queue for large diffs.
System architecture
01 Pull request Webhook event
02 Queue Large diffs batched
03 Context builder Diff + related code
04 Review agent Findings by severity
05 GitHub Inline comments
High-level flow of PRism, from input to output.
What you will learn
- Building and securing a GitHub App
- Giving an LLM repository context without blowing the context window
- Severity ranking and noise control for review comments
- Measuring reviewer usefulness from dismissals
What you will build
- Webhook service and job queue
- Context builder using syntax-aware chunking
- Inline comment poster with suggested fixes
- Usefulness dashboard
Build plan
- GitHub App and webhook
- Context builder
- Review agent with ranking
- Inline comments and suggestions
- Feedback loop and dashboard