Gemini CLI Skill for Claude Code — Summary

A single-skill Claude Code plugin that teaches Claude how to orchestrate Google's Gemini CLI as a powerful auxiliary AI tool. The skill provides detailed instructions, prompt templates, integration patterns, and reference docs so Claude can delegate tasks — code generation, security audits, test writing, web research, and architecture analysis — to Gemini 3 Pro running locally on the user's machine.

The core value proposition is multi-model parallelism without any backend infrastructure: Claude spawns gemini as a subprocess via Bash, the skill guides prompt construction and flag usage, and Claude post-processes the output. The --yolo / -o text pattern is the central primitive, with a JSON output variant for programmatic parsing.

The skill bundles five markdown files: SKILL.md (core instructions), reference.md (CLI flag reference), templates.md (reusable prompts), patterns.md (integration workflows like Generate-Review-Fix), and tools.md (Gemini-specific built-in tools including google_web_search, codebase_investigator, and save_memory).

This is a pure "skills-only behavioral framework" — no commands, no hooks, no MCP server. It extends Claude with awareness of a second AI rather than augmenting Claude's own reasoning. Compared to seeds: closest to superpowers (skills-only, zero commands, zero hooks), but entirely focused on cross-model delegation rather than spec-driven workflow enforcement. Unlike ccmemory (which uses Neo4j MCP for memory), this uses Gemini's native save_memory tool for cross-session persistence.

Gemini CLI Skill — Overview

Origin

Created by Foray Consulting (forayconsulting), published November 2025. No stated prior art or inspiration; the repo description is "A Claude Code skill enabling Claude to use Gemini 3 Pro via Gemini CLI." The README frames the skill as a bridge between two frontier LLMs, not as a workflow methodology.

Philosophy

The central claim: different AI models bring different "mindsets." Claude writing code benefits from a Gemini review pass precisely because it uses a different training data distribution and a different architecture. The skill instantiates this as the "Generate-Review-Fix Cycle" — Claude generates, Gemini critiques, Claude fixes.

A secondary claim: Gemini has unique tools (Google Search grounding, codebase_investigator, save_memory) that Claude Code cannot access natively. The skill is the bridge.

Manifesto-Style Quotes (from SKILL.md)

"When NOT to Use: Simple, quick tasks (overhead not worth it) ... Tasks requiring immediate response (rate limits cause delays) ... When context is already loaded and understood."

"YOLO Mode Behavior: Auto-approves tool calls but does NOT prevent planning prompts. Gemini may still present plans and ask 'Does this plan look good?' Use forceful language: 'Apply now' / 'Start immediately' / 'Do this without asking for confirmation'."

"Gemini's Unique Capabilities: These tools are available only through Gemini: google_web_search — Real-time internet search via Google; codebase_investigator — Deep architectural analysis; save_memory — Cross-session persistent memory."

Scope

5 files, no binary, no server. Installed by copying ~/.claude/skills/gemini-cli/ manually. Designed as a bolt-on capability layer, not a full development methodology.

Gemini CLI Skill — Architecture

Distribution

standalone-repo — a GitHub repo containing only markdown skill files. No npm package, no CLI, no hooks.

Install

git clone https://github.com/forayconsulting/gemini_cli_skill.git
cp -r gemini_cli_skill ~/.claude/skills/gemini-cli

Manual file copy. Claude Code auto-discovers skills in ~/.claude/skills/.

Prerequisites

• Gemini CLI installed: npm install -g @google/gemini-cli
• Gemini API key or OAuth authentication configured (free tier available)

Directory Tree

gemini_cli_skill/
├── SKILL.md         # Core skill definition — when/how to use Gemini CLI
├── reference.md     # Full CLI flag reference (--yolo, -o, -m, -r, etc.)
├── templates.md     # Reusable prompt templates (generation, review, test, docs)
├── patterns.md      # Integration patterns (Generate-Review-Fix, background execution)
└── tools.md         # Gemini built-in tools (google_web_search, codebase_investigator, save_memory)

Runtime Dependencies

• None server-side; Gemini CLI runs as a subprocess invoked via Claude's Bash tool
• gemini binary must be on PATH
• Free tier: 60 requests/min, 1000/day

Target AI Tools

• Primary: Claude Code (skill discovery and Bash invocation)
• Other tools that support skills (Gemini CLI itself has skill discovery, but this skill is for Claude to invoke Gemini, not the reverse)

Key Architectural Pattern

Claude Code reads SKILL.md into context when the use-case matches. Claude then constructs shell commands and invokes them through its native Bash tool. Gemini runs as an external process. Output comes back through stdout, parsed by Claude.

User → Claude Code
  → reads SKILL.md (into context)
  → Bash: gemini "[prompt]" --yolo -o text
    → Gemini CLI process (external, Google API call)
    → stdout response
  → Claude processes result
  → User gets answer

This is not an MCP server. There is no persistent connection. Each gemini invocation is a fresh subprocess.

Gemini CLI Skill — Components

Skills (1)

Supporting Reference Files (4)

Commands (0)

None.

Hooks (0)

None.

MCP Servers (0)

None. Gemini CLI is invoked directly via Bash subprocess, not via MCP protocol.

Scripts (0)

None.

Templates (0)

None in the template-as-file sense; templates.md contains prompt text, not file scaffolds.

Model Selection Logic (from SKILL.md)

• Default: gemini-2.5-pro (not specified = default)
• Fast/simple tasks: gemini -m gemini-2.5-flash
• The skill documents the tradeoff but leaves the choice to Claude's judgment

Gemini CLI Skill — Uniqueness

differs_from_seeds

Closest seed architecturally is superpowers (skills-only, zero commands, zero hooks, zero MCP). But where superpowers enforces a spec-first/TDD workflow methodology, the Gemini CLI skill enables a single capability: cross-model delegation to Gemini. The fundamental difference is purpose — superpowers changes how Claude reasons, this skill changes which model does the work. The secondary differentiator is that superpowers never routes work to external AI; it treats Claude as the sole reasoning engine and uses subagents only as parallel Claude instances. This skill makes Google's Gemini 3 Pro a first-class peer in the workflow.

Positioning

This is the only framework in the corpus (across all batches) whose primary value is explicit cross-model routing from Claude to Gemini. It occupies a unique niche: "multi-model coordination as a skill" rather than multi-model coordination as an orchestration layer (like Deebo or OpenRouter-based frameworks).

Distinctive Opinion

The framework implicitly argues that different LLMs catch different bugs — Gemini's security review catches things Claude misses, and vice versa. This is a testable but unvalidated claim in the README.

Observable Failure Modes

1. Rate limit surprise: Free tier (60 req/min, 1000/day) is hit in large projects; no graceful degradation built in
2. YOLO mode planning prompts: Despite --yolo, Gemini may still ask for confirmation; Claude must use forceful language or the workflow stalls
3. Output parsing ambiguity: JSON mode works, but text mode output varies; Claude must robustly handle Gemini's variable output formats
4. Context pollution: When Gemini generates long output (architecture analysis), the gemini stdout goes into Claude's context window, potentially bloating it
5. No feedback loop: Claude cannot programmatically evaluate Gemini's output quality; it must rely on its own judgment

What Makes It Novel vs. the 11 Seeds

• Only framework to treat another AI CLI as a tool rather than the user or Claude as agents
• Only framework with a documented multi-model model-selection heuristic (Pro vs Flash based on task complexity)
• Only skill that documents rate-limit strategies