gemini-peer-review

Gemini Peer Review

Consult Google's Gemini CLI to get a second opinion on code, architecture, debugging, or security. Synthesize both perspectives into a unified recommendation for the developer.

Invocation

/gemini-peer-review $ARGUMENTS

Task Classification

Read $ARGUMENTS and classify into one of four modes. If ambiguous, default to code-review.

Mode	Triggers	Gemini Model
code-review	"review", "check", "bugs", file paths without other context	gemini-2.5-flash
architecture	"architecture", "design", "scalability", "refactor", "pattern"	gemini-2.5-pro
debug	"debug", "error", "fix", "broken", "failing", stack traces	gemini-2.5-flash
security	"security", "audit", "vulnerability", "auth", "injection", "XSS"	gemini-2.5-pro

For tasks involving complex trade-off analysis or multi-system reasoning, upgrade to gemini-3-pro-preview regardless of mode.

Execution Steps

1. Classify the task from $ARGUMENTS using the table above.
2. Gather context — read relevant files, collect error messages, recent diffs (git diff), or architecture docs as needed.
3. Select the prompt template from references/prompt-templates.md matching the classified mode.
4. Construct the Gemini prompt using the AI-to-AI format from the template. Always prefix with:

   [Claude Code consulting Gemini for peer review]
   

   Fill in task-specific sections with gathered context.
5. Invoke Gemini non-interactively:

   # For file-based review (pipe content):
   cat <file> | gemini -m <model> -p "<prompt>"
   
   # For multi-file or general queries:
   gemini -m <model> -p "<prompt with inline context>"
   

6. Synthesize findings — present a unified analysis:
   • My analysis: your own findings from reading the code
   • Gemini's analysis: key points from Gemini's response
   • Unified recommendation: reconciled view, noting where you agree and any disagreements with reasoning

Prompt Construction

Use the templates in references/prompt-templates.md. The core format is always:

[Claude Code consulting Gemini for peer review]

Task: [Selected template for the classified mode, with context filled in]

Provide direct analysis with file:line references. I will synthesize your findings with mine before presenting to the developer.

Append gathered file contents, error messages, or diffs directly into the prompt body after the task description.

Model Selection Logic

Apply in order:

1. If $ARGUMENTS mentions a specific model name (flash, pro, 3-pro), use it.
2. If the task involves complex trade-offs across multiple systems → gemini-3-pro-preview
3. If mode is architecture or security → gemini-2.5-pro
4. Otherwise → gemini-2.5-flash

Output Format

Present findings to the developer as:

## Peer Review: [Mode] — [brief subject]

### My Analysis
[Your findings with file:line references]

### Gemini's Analysis
[Key findings from Gemini, attributed]

### Unified Recommendation
[Reconciled view — agreements, disagreements with reasoning, action items]

Error Handling

• If gemini CLI is not found: tell the developer to install it (npm install -g @anthropic-ai/gemini-cli or check Google's docs) and retry.
• If Gemini returns an error or empty response: report the error, present your own analysis only, and note Gemini was unavailable.
• If context is too large for a single prompt: split into focused sub-queries (e.g., review file-by-file) and aggregate.

Constraints

• Never forward raw Gemini output without synthesis.
• Always attribute which findings come from Gemini vs your own analysis.
• Do not pipe sensitive credentials or environment variables to Gemini.
• Respect .gitignore patterns when gathering file context.