Code Review Prompts

This directory contains prompt templates used by the AI code reviewer to analyze pull request changes.

Available Prompts

File	Description
generic-code-review.md	General code review prompt applicable to all languages
languages/csharp-code-review.md	C# and .NET specific patterns and anti-patterns
languages/typescript-code-review.md	TypeScript/JavaScript and React specific patterns

How Prompts Work

The code reviewer selects prompts based on the file extensions in the PR:

• .cs files → Uses languages/csharp-code-review.md
• .ts, .tsx, .js, .jsx files → Uses languages/typescript-code-review.md
• Other files → Uses generic-code-review.md

Customizing Prompts

Adding a New Language

1. Create a new file: {language}-code-review.md
2. Follow the structure of existing prompts:
   • Language-specific focus areas
   • Common anti-patterns with examples
   • Good patterns to recognize
   • Output format specification
3. Update PromptService.cs to use the new prompt for relevant file extensions

Modifying Existing Prompts

You can customize prompts to match your team's coding standards:

1. Add company-specific rules: Include internal best practices
2. Adjust severity levels: Change what constitutes Critical vs Minor
3. Add/remove categories: Focus on areas important to your codebase
4. Include framework-specific guidance: Add rules for your tech stack

Example: Adding Company Rules

## Company-Specific Rules

### Logging
- All public methods in services must have trace logging
- Error logs must include correlation ID
- Never log PII (personally identifiable information)

### Naming Conventions
- Repository classes must end with `Repository`
- Service classes must end with `Service`
- DTOs must end with `Dto` or `Request`/`Response`

Prompt Structure

Each prompt should contain:

1. Role Definition

Tell the AI what expertise it should apply.

2. Focus Areas

List specific things to look for, organized by category.

3. Severity Definitions

Define what each severity level means for your team.

4. Anti-Pattern Examples

Show concrete code examples of what to flag.

5. Good Pattern Examples

Show what correct code looks like.

6. Output Format

Specify the exact JSON structure expected.

Testing Prompts

To test prompt changes:

1. Run the function locally
2. Create a test PR with known issues
3. Verify the AI catches expected problems
4. Adjust prompt wording as needed

Tips for Effective Prompts

1. Be Specific: Vague instructions lead to vague reviews
2. Show Examples: The AI learns from examples better than rules
3. Balance Coverage: Too many rules can dilute focus
4. Test Iteratively: Small changes, test, repeat
5. Consider Context: Remember the AI sees code in isolation

Prompt Token Limits

Keep prompts concise to leave room for code:

• System prompt: ~1,000-2,000 tokens recommended
• User prompt (code): Up to 100,000 tokens with gpt-4o
• Code per request: Limited by MaxLinesPerRequest setting (default: 2,000 lines)

Large prompts reduce the amount of code that can be reviewed in a single request.