A walkthrough of building an AI-powered code review agent in C# that uses Git diffs as input to LLMs instead of entire repositories. Covers the five-step workflow: extracting diffs via Git, constructing focused prompts, sending requests to LLMs (OpenAI, Azure OpenAI, Ollama, Claude, Gemini), parsing structured JSON responses, and surfacing feedback in tools like VS Code, GitHub PRs, or Azure DevOps. Also discusses best practices such as limiting token usage, caching repeated reviews, adding severity levels, and advanced enhancements like multi-agent reviews, semantic context retrieval, and local LLM support for privacy.
Table of contents
Why Git Diff Is the Right ApproachHigh-Level ArchitectureSolution FlowStep 1 - Extract Git Diff in C#Step 2 - Build the AI PromptStep 3 - Send Request to LLMStep 4 - Parse AI ResponseStep 5 - Show Feedback to DevelopersAdding Smart ContextBest Practices for AI Code ReviewMulti-Agent ReviewsLocal LLM SupportPull Request AutomationCommon ChallengesSensitive Code ExposureFinal Thoughts658 Impressions