gemini-rescue
Hand off the current problem to Google Gemini for an independent read. Two modes — pick the one that matches the task.
When to use
- Adversarial review: Claude has a plan or implementation that looks right but is high-blast-radius (auth, migrations, payments, infra). Get an uncorrelated reviewer.
- Stuck: Claude has tried 2+ angles and is going in circles. A second vendor often sees a different path.
- 1M-context audit: The codebase or document is larger than fits in Claude's context. Gemini 2.5 Pro handles 1M tokens natively.
- NOT for: simple lookups (use Claude directly), Gemini-specific features (call the SDK directly), or tasks where you already know Gemini-2.5-Flash won't add signal.
Two modes
Mode A — One-shot review (gemini_ask)
For diff or plan review, classification, scoring, structured output. Fast, cheap, stateless. Claude stays the orchestrator.
Call the gemini_ask MCP tool with:
prompt: the full review prompt (include the diff/plan/problem inline)system_prompt: role framing (e.g., "You are an adversarial code reviewer. Find what's wrong, not what's right.")model: optional override (default is set in the MCP server env)temperature: lower for review, higher for brainstorming
Treat the response as one opinion, not ground truth. Cross-reference with the actual code.
Mode B — Agentic dispatch (gemini_run)
For tasks where you want Gemini to do the work itself — read files, navigate the repo, propose patches. Slower, more expensive, but you get a full second agent.
Call the gemini_run MCP tool with:
task: the full task description (include any file paths Gemini should read)cwd: working directory Gemini should operate intimeout_seconds: hard kill (default 300)output_format: text or json
Use this for 1M-context audits, parallel agent-on-agent work, or when you'd otherwise spawn a Codex rescue.
After the rescue
- Read Gemini's response critically — it can hallucinate, miss context, or just be wrong.
- Verify any specific claims (function names, file paths, model IDs) against the actual code.
- If Gemini surfaced something Claude missed, name what it caught and why Claude missed it — that's the value of vendor orthogonality.
- Do NOT auto-apply Gemini's patches. Treat them as suggestions, the same way you'd treat a teammate's review.
Prerequisites
This skill requires a Gemini MCP server registered in your Claude Code settings. The MCP server should expose gemini_ask and gemini_run tools. Configure it with your Google API key.
Related
/codex:rescue — same idea, OpenAI as the second vendor. Use Codex for code-specific problems where GPT excels; use Gemini for long-context, multimodal, or when Codex has already been tried.
1---2name: gemini-rescue3description: Get a second opinion from Google Gemini on the current diff, plan, or problem. Use when Claude is stuck, when an implementation needs adversarial review by a different vendor, or when a 1M-context audit is needed that Claude can't fit. Trigger on "second opinion", "rescue with Gemini", "let Gemini review this", "adversarial review", "1M-context audit", "/gemini-rescue".4---56# gemini-rescue78Hand off the current problem to Google Gemini for an independent read. Two modes — pick the one that matches the task.910## When to use1112- **Adversarial review**: Claude has a plan or implementation that looks right but is high-blast-radius (auth, migrations, payments, infra). Get an uncorrelated reviewer.13- **Stuck**: Claude has tried 2+ angles and is going in circles. A second vendor often sees a different path.14- **1M-context audit**: The codebase or document is larger than fits in Claude's context. Gemini 2.5 Pro handles 1M tokens natively.15- **NOT for**: simple lookups (use Claude directly), Gemini-specific features (call the SDK directly), or tasks where you already know Gemini-2.5-Flash won't add signal.1617## Two modes1819### Mode A — One-shot review (`gemini_ask`)2021For diff or plan review, classification, scoring, structured output. Fast, cheap, stateless. **Claude stays the orchestrator.**2223Call the `gemini_ask` MCP tool with:24- `prompt`: the full review prompt (include the diff/plan/problem inline)25- `system_prompt`: role framing (e.g., "You are an adversarial code reviewer. Find what's wrong, not what's right.")26- `model`: optional override (default is set in the MCP server env)27- `temperature`: lower for review, higher for brainstorming2829Treat the response as one opinion, not ground truth. Cross-reference with the actual code.3031### Mode B — Agentic dispatch (`gemini_run`)3233For tasks where you want Gemini to *do the work itself* — read files, navigate the repo, propose patches. Slower, more expensive, but you get a full second agent.3435Call the `gemini_run` MCP tool with:36- `task`: the full task description (include any file paths Gemini should read)37- `cwd`: working directory Gemini should operate in38- `timeout_seconds`: hard kill (default 300)39- `output_format`: `text` or `json`4041Use this for 1M-context audits, parallel agent-on-agent work, or when you'd otherwise spawn a Codex rescue.4243## After the rescue44451. Read Gemini's response critically — it can hallucinate, miss context, or just be wrong.462. Verify any specific claims (function names, file paths, model IDs) against the actual code.473. If Gemini surfaced something Claude missed, name what it caught and why Claude missed it — that's the value of vendor orthogonality.484. Do NOT auto-apply Gemini's patches. Treat them as suggestions, the same way you'd treat a teammate's review.4950## Prerequisites5152This skill requires a Gemini MCP server registered in your Claude Code settings. The MCP server should expose `gemini_ask` and `gemini_run` tools. Configure it with your Google API key.5354## Related5556- `/codex:rescue` — same idea, OpenAI as the second vendor. Use Codex for code-specific problems where GPT excels; use Gemini for long-context, multimodal, or when Codex has already been tried.
m2ai-portfolio