gemini

---
name: gemini
description: "Gemini inside Claude Code. Second opinions, web-grounded facts, image generation, code visualization, and multi-round brainstorms — all via one skill. Use when user says 'ask Gemini', 'second opinion', 'cross-validate', 'check with Gemini', 'ground this', 'verify this claim', 'generate image', 'visualize', 'draw architecture', 'diagram', 'brainstorm with Gemini', or needs an independent opinion from a different model family. Also trigger when evaluating hypotheses, fact-checking, generating diagrams from code, or wanting adversarial multi-model ideation."
allowed-tools: Bash, Read, Glob, Grep, mcp__gemini-sdk__gemini_image
---

# Gemini Skill — Full Toolkit for Claude Code

Different model family = different blind spots. This skill gives Claude direct access to Gemini for text, search, images, and multi-round dialogue.

## Setup (first run)

1. Get API key at https://aistudio.google.com → create key
2. Add to `.env`: `GOOGLE_API_KEY=your_key`
3. Install SDK: `pip install google-genai`
4. Find `gemini.py`: bundled at `scripts/gemini.py` relative to this SKILL.md

```bash
GEMINI="$(find . ~/.claude/skills -name gemini.py -path '*/gemini/*' 2>/dev/null | head -1)"
```

## 1. Text — Ask, Think, Review

```bash
# Quick question (Flash — fast, cheap)
python3 $GEMINI ask "What's the difference between REST and GraphQL?"

# Second opinion (Pro — deep reasoning, critical)
python3 $GEMINI second-opinion "Is this the right architecture?" --context "context here"

# Deep reasoning (Pro — maximum thinking)
python3 $GEMINI think "Design a caching strategy for 10M users"

# Code review
python3 $GEMINI review "def process(data): ..." --focus "error handling"

# Data analysis
python3 $GEMINI analyze "revenue data shows..." --context "Q1 report"

# Structured extraction
python3 $GEMINI extract "Parse this invoice..." --json-mode
```

| Command | Default Model | System Instruction | Use When |
|---------|---------------|-------------------|----------|
| `ask` | 3-flash-preview | none | Quick questions, general tasks |
| `second-opinion` | **3.1-pro-preview** | Critical reviewer | Decisions, validation, critical review |
| `think` | **3.1-pro-preview** (high thinking) | none | Complex reasoning, architecture |
| `review` | 3-flash-preview | Code reviewer | Code quality, bugs |
| `analyze` | 3-flash-preview | Data analyst | Data patterns, CSV analysis |
| `extract` | 3-flash-preview | JSON extractor | Structured JSON from text |

## 2. Web-Grounded Answers

Add `--grounded` to any command. Gemini searches Google before responding — real-time facts, not training data.

```bash
# Current version of a library
python3 $GEMINI ask "What's the latest Next.js version?" --grounded

# Verify a claim with sources
python3 $GEMINI second-opinion "Is Clerk still free for 10K MAU?" --grounded

# Research mode (Flash-Lite — cheapest, fastest for facts)
python3 $GEMINI ask "Compare Drizzle vs Prisma pricing 2026" \
  -m gemini-3.1-flash-lite-preview --grounded --save research.md
```

## 3. Multimodal — Images as Input

Use `--image` / `-i` (repeatable) to send images with any command:

```bash
# Describe a screenshot
python3 $GEMINI ask "Describe this UI" --image screenshot.png

# Compare designs (multiple images)
python3 $GEMINI second-opinion "Compare designs" --image site.png --image ref.png --save review.md

# Visual QA with file prompt
python3 $GEMINI second-opinion @prompt.txt --image site.png --image reference.png --save review.md
```

Supported: PNG, JPEG, WebP, GIF. Max ~20MB total. Gemini 3.x is natively multimodal.
Use cases: design review, visual QA, screenshot diff, accessibility audit.

## 4. Image Generation

Generate images via Gemini Nano Banana models. Requires `mcp__gemini-sdk__gemini_image` MCP tool.

```
# From Claude Code — just describe what you want:
"Generate an architecture diagram for my project"
"Create a banner image with dark background and gold accents"
"Draw a flow chart showing the auth process"
```

The skill uses Gemini Flash Image (fast) or Pro Image (quality). Supports 14 aspect ratios, 512-4K resolution.

For **code-to-diagram** generation, see the Visualize workflow below.

## 5. Visualize — Code to Diagrams

Turn any codebase into professional diagrams. Three phases: Gather → Synthesize → Render.

**Types:** `system-map` | `flow` | `connections` | `data-model`

**Quick usage:**
```
"Visualize the architecture of this project"
"Draw how the API pipeline works"
"Show me the data model"
"How are these services connected?"
```

**Process:**
1. **Gather** — read interfaces, configs, entry points (max 10 files, never .env)
2. **Synthesize** — convert to Nodes + Edges in plain English (anti-hallucination: Gemini never sees raw code)
3. **Render** — call Gemini Image with 1500+ char prompt built from synthesis

For detailed gathering strategies and prompt templates, read `references/visualize.md`.

## 6. Brainstorm — 3-Round Claude x Gemini

Structured adversarial dialogue. Flash researches facts, Pro reasons on them, Claude orchestrates.

```
"Brainstorm how to launch this product"
"Let's think through options for the architecture"
"Diverge and converge on the best approach"
```

**Pipeline:** Ground (dual web-check) → R1: Diverge (7+ ideas) → R2: Deepen (kill to 3) → User check-in → R3: Converge (1 winner) → Fact-check → One decision.

For the full protocol with prompt templates, read `references/brainstorm.md`.

## Arguments

| Arg | Short | Description |
|-----|-------|-------------|
| `--model` | `-m` | Override model (e.g., `gemini-3.1-flash-lite-preview`) |
| `--context` | `-c` | Additional context (second-opinion, analyze) |
| `--grounded` | `-g` | Enable Google Search grounding — Gemini searches the web before answering |
| `--save` | | Save response to file |
| `--thinking` | `-t` | Level: minimal/low/medium/high (Gemini 3.x) |
| `--json-mode` | | Force structured JSON output |
| `--image` | `-i` | Attach image file (repeatable for multiple images) |
| `--system` | `-s` | Override system instruction |
| `--temp` | | Temperature: 0.0-2.0 (default 1.0, **keep default for Gemini 3**) |
| `--top-p` | | Top-p nucleus sampling: 0.0-1.0 (default 0.95) |
| `--top-k` | | Top-k sampling (default ~40) |
| `--max-tokens` | | Max output tokens: 1-65536 (default model max 64K) |
| `--seed` | | Seed for reproducible output |
| `--focus` | | Focus area for review |
| `--json` | | Raw JSON output (for piping) |
| `@file` | | Read prompt from file instead of inline |

## Key Parameter Rules (Gemini 3)

- **Temperature 1.0 is mandatory** — Google strongly recommends keeping default. Lower values cause looping/degradation in reasoning tasks.
- **thinking_level replaces thinking_budget** — Do NOT combine both (→ error). Pro: low/medium/high. Flash: minimal/low/medium/high.
- **JSON mode** — Use `--json-mode` for guaranteed valid JSON output.
- **Grounding is a Tool** — `--grounded` adds Google Search as a tool, not a parameter.

## Models (March 2026)

| Model | Price (in/out per 1M) | Best for |
|-------|----------------------|----------|
| `gemini-3.1-pro-preview` | $2 / $12 | Reasoning, brainstorm, second opinions |
| `gemini-3-flash-preview` | $0.50 / $3 | General tasks, code review (default) |
| `gemini-3.1-flash-lite-preview` | $0.25 / $1.50 | Research, grounding, fact-check |

For full model specs, read `references/models.md`.

## Critical Evaluation Rule

Gemini's output is INPUT for decisions, not the decision itself. After every call:
1. **Challenge each recommendation** — is this fact or speculation?
2. **Check for missing context** — Gemini doesn't see codebase, prior decisions, or constraints
3. **Verify numbers** — predictions like "90% speedup" are estimates, not measurements
4. **Look for blind spots** — implementation complexity, side effects, existing code
5. **Present critical assessment to user** — "Gemini said / My evaluation / Reason" table

**Workflow:** Gemini recommends → Claude evaluates critically → present BOTH to user → user decides

## Parallel Execution

```bash
python3 $GEMINI second-opinion @aspect1.txt --save r1.md &
python3 $GEMINI second-opinion @aspect2.txt --save r2.md &
python3 $GEMINI second-opinion @aspect3.txt --save r3.md &
wait
```

## References

Read these only when needed:
- `references/models.md` — all models, pricing, thinking levels
- `references/parameters.md` — complete API parameter reference
- `references/visualize.md` — diagram generation guide (types, gathering, synthesis, rendering)
- `references/brainstorm.md` — 3-round protocol with prompt templates
- `references/decision-framework.md` — when to use Gemini vs Claude subagents