Gemini inside Claude Code. Second opinions, web-grounded facts, diagrams, brainstorms — one skill.
Different model family = different blind spots. Use both.
Claude is great. But a single model has blind spots. Gemini sees differently — different training, different reasoning patterns, different knowledge cutoffs. This skill gives Claude direct access to Gemini for everything: quick questions, critical reviews, real-time web facts, image generation, code visualization, and structured multi-round brainstorms.
One install. Five capabilities.
Claude Code (recommended):
/plugin marketplace add awrshift/skill-gemini
Manual:
mkdir -p .claude/skills/gemini/scripts .claude/skills/gemini/references
curl -sL https://raw.githubusercontent.com/awrshift/skill-gemini/main/skills/gemini/SKILL.md \
-o .claude/skills/gemini/SKILL.md
curl -sL https://raw.githubusercontent.com/awrshift/skill-gemini/main/skills/gemini/scripts/gemini.py \
-o .claude/skills/gemini/scripts/gemini.py- Get API key at aistudio.google.com → Create API Key
- Add to
.env:GOOGLE_API_KEY=your_key_here - Install:
pip install google-genai
That's it. The skill auto-discovers gemini.py from its bundle.
| You say | What happens |
|---|---|
| "Ask Gemini about X" | Quick answer via Flash (fast, cheap) |
| "Get a second opinion" | Critical review via Pro (deep reasoning) |
| "Think deeply about X" | Maximum reasoning via Pro (high thinking) |
| "Review this code" | Code review with focus areas |
python3 gemini.py ask "Quick question"
python3 gemini.py second-opinion "Is this right?" --context "details"
python3 gemini.py think "Design a caching strategy"
python3 gemini.py review "def process(data): ..." --focus "error handling"Gemini searches Google before responding. Real-time facts, not training data.
python3 gemini.py ask "Latest Next.js version?" --grounded
python3 gemini.py second-opinion "Is Clerk still free?" --groundedBanners, diagrams, illustrations — via Gemini Nano Banana models.
"Generate a dark banner for my GitHub repo"
"Create an architecture diagram"
Three-phase pipeline: read your codebase → synthesize nodes/edges → render via Gemini Image.
"Visualize the architecture of this project"
"Draw the data flow"
"Show me how these services connect"
Supports: system-map, flow, connections, data-model. Full guide in references/visualize.md.
Claude orchestrates. Flash researches facts. Pro reasons on them. 7 ideas → 3 survivors → 1 winner.
"Brainstorm how to launch this"
"Let's think through options with Gemini"
Full protocol with prompt templates in references/brainstorm.md. Cost: ~$0.35 per brainstorm.
| Model | Price (in/out) | Best for |
|---|---|---|
Pro (gemini-3.1-pro-preview) |
$2 / $12 per 1M | Reasoning, brainstorm, second opinions |
Flash (gemini-3-flash-preview) |
$0.50 / $3 per 1M | General tasks, code review (default) |
Flash-Lite (gemini-3.1-flash-lite-preview) |
$0.25 / $1.50 per 1M | Research, grounding, fact-check |
google-genainotgoogle-generativeai— the new SDK package name (2026)- Temperature 1.0 is mandatory — lower values cause looping in Gemini 3
--groundedis a tool, not a parameter — adds Google Search capability- Image generation requires Gemini MCP server —
mcp__gemini-sdk__gemini_image - Second opinion ≠ final answer — Claude should always evaluate Gemini's output critically
| Platform | Install |
|---|---|
| Claude Code | /plugin marketplace add awrshift/skill-gemini |
| Codex CLI | Copy skills/gemini/ to .openai/skills/ |
| Gemini CLI | Copy skills/gemini/ to .gemini/skills/ |
| Any Agent Skills tool | Follow spec |
- skill-awrshift — adaptive decision-making framework
- skill-brainstorm — standalone 3-round brainstorm (also bundled here)
- skill-telegram — Telegram user API integration
- ClawClaw Soul — persistent identity protocol for AI agents
- Fork the repository
- Create a feature branch
- Submit a pull request
MIT — see LICENSE for details.