Roast Review Skill

This skill performs a code review by asking Gemini to "roast" the codebase, then translating the roast into actionable feedback.

Why This Works

When asked to "roast" code, LLMs surface the most glaring issues - the things that would stand out to any developer. This is a useful heuristic for finding high-priority problems.

Configuration

Supported Models

Repomix Caching

The script caches repomix output in {tempdir}/repomix-{hash}.txt. Use --no-cache to regenerate.

Execution Steps

Step 1: Determine Scope

Ask the user what to roast:

• Entire repository (default)
• Specific directory (e.g., src/, tests/)

Step 2: Run the Roast

# Set target and output paths
TARGET_DIR="$(pwd)"  # or "$(pwd)/src" for specific directory
ROAST_OUTPUT="/tmp/roast-$(echo -n "$TARGET_DIR" | md5sum | cut -c1-12).md"

# Run the roast (use --repo for subdirectory)
python3 ${CLAUDE_SKILL_DIR}/scripts/repo_query.py --quiet --model gemini-3-pro-preview --output "$ROAST_OUTPUT" \
  "Roast this repo. Be brutally honest about what's wrong. For EVERY issue, include the specific file path and line number(s) in the format 'file.py:123' or 'file.py:100-150'. Point out bad practices, code smells, missing features, poor organization, security issues, and anything else problematic."

Step 3: Report Token Usage

The script outputs token usage and cost. Report this to the user:

Token usage: X input, Y output, Z total
Estimated cost: $0.XXXX (model-name)

Step 4: Read and Present Results

1. Use Read tool to get the roast from $ROAST_OUTPUT
2. Present the roast to the user
3. Translate each issue into constructive feedback with this format:

## Code Review Summary

### Critical Issues
- **[file:line]** Issue description → Recommendation

### High Priority
- **[file:line]** Issue description → Recommendation

### Medium Priority
- **[file:line]** Issue description → Recommendation

### Low Priority
- **[file:line]** Issue description → Recommendation

### Positive Observations
- What's good about the codebase

For each roasted issue:

• Extract the file:line reference
• Rewrite harshly-worded criticism as constructive feedback
• Add specific, actionable recommendation
• Prioritize by severity

Step 5: Validate Flagged Issues

IMPORTANT: Gemini may hallucinate file paths, line numbers, or issues that don't exist. You MUST validate each flagged issue.

For each issue with a file:line reference:

1. Use the Read tool to read the flagged file at the specified line(s)
2. Verify the issue actually exists in the code
3. Mark each issue with a validation status:

## Validated Issues

### Critical Issues
- **[file:line]** ✅ VERIFIED - Issue description → Recommendation
- **[file:line]** ❌ NOT FOUND - Gemini claimed X but the code shows Y

### Unverified Issues
- Issues where file/line couldn't be located or code differs significantly

Validation criteria:

• ✅ VERIFIED: The code at that location clearly shows the described issue
• ⚠️ PARTIAL: The issue exists but at a different line or in modified form
• ❌ NOT FOUND: File doesn't exist, line is out of range, or code doesn't match the claim
• 🤔 SUBJECTIVE: The "issue" is a style/opinion matter, not an objective problem

If more than 30% of issues fail validation, warn the user that the roast may be unreliable and suggest re-running with a different model.

Examples

User: "Roast this repo" Action: Run on current directory, read output, present with recommendations

User: "Roast the src directory" Action: Run with --repo src/, read output, present with recommendations

User: "Quick roast" Action: Use --model gemini-2.5-flash for faster results

Requirements

• GEMINIAPIKEY environment variable
• google-genai: pip install google-genai
• repomix: npm install -g repomix (or via npx)