Imagine

A second brain for your coding agent — Google Gemini inside Claude Code & Codex.
**Hand the hard thinking — reasoning, independent code review, deep research, and *automatic prompt-engineering* — to Google Gemini. It runs on your Google AI Pro / Ultra subscription (agy), *not* on your agent's tokens.**
⚡ One install · 4 skills · **Claude Code & Codex** · no API key · no extra bill — just your existing AI Pro plan.
> ⚠️ **Security:** every gemini call sends your prompt + piped files to Google verbatim. **Never pipe secrets** (.env, keys, tokens). **Never run Gemini's output unread** — piped files can carry prompt injection, and Gemini can be wrong. See **SECURITY.md**.
---
Why Imagine
Your coding agent is brilliant — and it has three blind spots:
- 🪞 **It reviews its own code.** Same model, same blind spots. Before you merge, you want a genuinely *different* pair of eyes.
- 🔥 **Research burns your tokens.** Digesting a huge module or comparing five libraries chews through your agent's budget fast.
- 🎲 **It acts on whatever you typed.** A rushed, half-formed prompt → a rushed, half-formed result.
**Imagine fixes all three** by giving your agent a second, *independent* model — **Google Gemini** — to hand work to. And because generation runs through Google's **Antigravity CLI** on your **AI Pro / Ultra subscription**, that heavy thinking is effectively **free**: a flat monthly fee you already pay, never metered against your agent's tokens.
> **A second brain for your coding agent — that you've already paid for.**
✨ The four skills
| Skill | What it gives you | Just say… |
|-------|-------------------|-----------|
| 🧠 **gemini-pro** | One-shot delegation to Gemini (default **3.1 Pro, High**), with model selection via -m. Offload hard reasoning and save your agent's credits. | *"ask Gemini Pro to…"*, *"use Gemini"*, *"юзай Gemini"* |
| 🔍 **gemini-review** | An **independent, adversarial** review of code / a diff / a PR / a plan — a different model, different blind spots, severity-tagged findings. Catch what self-review can't. | *"let Gemini review this"*, *"second opinion before I merge"* |
| 📚 **gemini-research** | **Deep research & large-context digestion** — summaries, comparisons, "find X across all of this", sourced reports — without spending agent tokens on the reading. | *"have Gemini dig into this module"*, *"compare these options"* |
| ✨ **gemini-proxy** | **Automatic prompt-engineering.** For your next N prompts, Gemini quietly rewrites each message into a sharper version *before* your agent acts. | *"enable gemini proxy"*, *"polish my prompts"*, *"переделывай мои промты через гемини"* |
Plus **bin/gemini** — a dependency-free Bash wrapper over agy -p (model shortcuts + stdin context), and **install.sh** to wire it all up for **both** Claude Code and Codex.
⭐ The one people love: gemini-proxy
The other three skills are one-shot. **gemini-proxy turns Gemini into your agent's built-in prompt engineer.** Enable it, and for your **next N substantive prompts (default 10)** Gemini rewrites each message you type into a sharper, more detailed **first-person** version of *itself* — as a senior prompt engineer would — *before* your agent acts. Your agent still answers **with the full chat context, your project rules, and memory**; Gemini only sharpens the wording. A **smart filter** routes real work (build / fix / review / design) and skips trivial / config / meta chatter, so the counter only spends on substance. It **self-disables after N**.
**Ground it first — this is what makes it great.** Drop a rules file in your project (canon + safety invariants + standing expectations) and Gemini rewrites *with your project in mind*: