Gemini Skill for Claude Code

Ground your Claude Code work in reality with Google Gemini's 1M token context

  

   

---

Cross-check your Claude Code work with Google Gemini. Search current docs, validate against best practices, ingest entire codebases. Gemini's strength is grounding — connecting your code to the real world.

Install | How It Works | Modes | Grounding | Contributing

Why This Exists

Claude is great at code. Codex is great at adversarial review. But sometimes you need something different — you need to know if your approach matches current best practices, if the library you're using has known issues, or if there's a better pattern the ecosystem has converged on.

Gemini does this. It has a 1M token context window (ingest your entire codebase), web search built in (check current docs and GitHub issues), and a different reasoning style that triangulates answers you can't get from Claude or Codex alone.

Install

# Copy to your Claude Code skills directory
mkdir -p ~/.claude/skills/gemini
curl -sL https://raw.githubusercontent.com/199-biotechnologies/gemini-skill/main/SKILL.md \
  -o ~/.claude/skills/gemini/SKILL.md

Requirements:

• Gemini CLI installed
• Google OAuth authenticated (run gemini interactively once)

How It Works

┌─────────────┐     prompt + context     ┌─────────────┐
│  Claude Code │ ───────────────────────► │  Gemini CLI  │
│  (you work   │                          │  Auto-routes │
│   here)      │ ◄─────────────────────── │  + fallback  │
└─────────────┘     output .md file      └─────────────┘

Claude writes a structured prompt, runs gemini with auto-routing (no forced model — handles 429s with automatic fallback), and reads the output file back. Free tier available.

Key design choice: The skill never forces a model with -m. Auto-routing picks the best available model and falls back automatically on rate limits. This is more reliable than pinning a specific model.

Modes

Mode	What Happens	When to Use
Quick	Inline question, direct answer	"What does Gemini think about X?"
Analysis	Claude curates context, Gemini analyzes	Architecture validation, broad analysis
Delegated	Gemini explores codebase + searches web	Grounding, research, ecosystem checks

Quick Mode

"Ask gemini if this API design follows REST conventions"

Fast, inline. Gemini answers from its training data.

Analysis Mode

"Get gemini's take on our database schema"

Claude summarizes what it knows. Gemini analyzes the curated context. No filesystem access needed.

Delegated Mode

"Have gemini research current best practices for WebSocket auth"

Gemini gets full filesystem access AND web search. It reads your code, searches for current docs, checks GitHub issues, and grounds its analysis in real-world state.

Grounding

Grounding is Gemini's killer feature in this skill. When you ask Gemini to ground something, it:

1. Reads your code to understand what you built
2. Searches the web for current docs, best practices, CVEs
3. Checks GitHub for related issues, patterns, alternatives
4. Compares your approach against what the ecosystem recommends
5. Reports with evidence, not just opinions

"Have gemini ground our auth implementation against current OWASP guidelines"

When to Use Gemini vs Codex

Task	Gemini	Codex
"Is our approach standard?"	Yes — web search	No
"Are we using this library correctly?"	Yes — checks current docs	Maybe
"Find security bugs"	Either	Preferred
"What does the ecosystem look like?"	Yes — broad research	No
"Implement this feature"	Either	Preferred

Configuration

• Auto-routing — no model forced, automatic 429 fallback
• Maximum thinking depth — thinkingLevel: HIGH by default
• YOLO mode — auto-approves all tool calls for non-interactive use
• Output persistence — saved to ~/.claude/subagent-results/gemini-output-*.md

Trigger Phrases

Say any of these in Claude Code to activate the skill:

• "use gemini", "ask gemini", "gemini check"
• "double check with gemini", "verify with gemini"
• "gemini review", "gemini implement"
• "what does Gemini think", "Google's take"
• "ground this", "cross-check with gemini"

Contributing

PRs welcome. If you've found better prompting patterns for grounding or a workflow improvement, open a PR.

License

MIT

---

Built by Boris Djordjevic at Paperfoot AI

If this is useful to you: