One model has the eye. The other keeps the wheel. ask-gemini puts them in the same car — and lets you stop driving with your knees.
ask-gemini is a Claude Code skill that hands your UI/UX work to a second model — the one with actual taste — while Claude stays in charge of everything that can break. You get the design talent of one AI and the steady hands of another, from a single command, without ever leaving your editor.
It's not a chatbot bolted on the side. It's delegation.
You know the trade you keep making.
Here's the thing nobody says out loud: Gemini has the best UI taste of the bunch. Guided tightly and precisely, it nails visual design, holds consistency across screens better than the Claude models, and does it faster. That's the upside.
The downside is just as real. Let Gemini actually build — touch the code, own the frontend, run loose in a real codebase — and it falls apart. Wrong imports, half-wired components, things quietly snapping. On its own, it's unusable for engineering.
So you pick your poison: settle for so-so UI from a model that can code, or babysit the one with taste by hand, re-explaining your project every time you switch tools. Either way, Gemini's best quality sits there unused — because nobody trusts it to drive.
The fix isn't to trust it more. It's to never let it drive. Leverage its taste through Claude — the main agent holds the wheel, Gemini just paints. ask-gemini is that arrangement, made automatic.
Think of it as a two-person team:
- Gemini is the eye. Design, visual judgment, UX instinct. That's the only job it gets.
- Claude is the hand on the wheel. Integration, wiring, cleanup — and it reviews everything Gemini sends back before a single line touches your code.
Each model does the one thing it's great at. Frontend work speeds up because nobody's fighting their own weakness.
(Under the hood it can run Gemini, Claude, or other models — but "ask Gemini" is the everyday move, so that's the name.)
- You ask — "redesign the watchlist card."
- Claude fires one command and hands the task off.
- Gemini does the design work — and sends it back.
- Claude reviews, adapts, and merges it into your actual project.
- You get the result — plus a note on what changed and why.
No copy-paste. No re-explaining. The handoff is invisible.
- A second opinion on a design or a decision
- Building UI — components, layouts, color, whole screens
- A critique of a screen that feels off
- Find-and-fix — it spots the problems and drafts the repairs
- Offloading the frontend entirely while Claude keeps grinding the backend
| Feature | What it gets you |
|---|---|
| Topic memory | Name a screen once ("watchlist"). Come back days later — it remembers. Refine screen after screen and the whole app ends up looking like one person designed it. |
| Chains | Bigger jobs run as steps that build on each other — ideate 5 → pick the best → build it — each step remembering the last, so the build never drifts from the idea. |
| You set the leash | Want critique only? It suggests. Want fixes too? It ships them. Your call. |
| Two lanes at once | Gemini takes the frontend while Claude takes the backend — in parallel. Less waiting, less cost, nothing broken. |
| Background mode | Long job? It works while you keep moving and pings you when it's done. |
| Boring on purpose | Clean predictable output, never cuts off a model mid-thought, full test suite underneath. Boring plumbing is exactly what you want when two models share a project. |
"Gemini breaks things — why let it near my code?"
Because it never touches the wheel. Claude reviews everything it sends back before a line lands — so the failure mode that makes Gemini risky solo never gets the chance. You get the taste; Claude keeps the safety.
Just talk. Ask for design help, a second opinion, or say "ask gemini" — the skill triggers on its own. No commands to memorize.
One-time setup:
# 1. install the skill globally (every agent on this machine)
npx skills add -g ASACHIT/ask-gemini
# 2. wire up the `ask-gemini` command + check the agy engine
bash ~/.agents/skills/ask-gemini/scripts/install.shNeeds the Antigravity CLI (agy) installed and signed in once (run agy once interactively) — you install agy yourself; the setup script only checks it's there. Works in any tool on the open Agent Skills standard — Claude Code, OpenAI Codex CLI, Pi CLI, GitHub Copilot CLI, Antigravity/Gemini.
Here's where it gets fun. ask-gemini doesn't run alone — pair it with any other skill and tell it to use it. That skill runs inside Gemini while it works, so Gemini designs with your design system, your linter, your component rules already in hand. Point a frontend skill at it and the output comes back already shaped to your conventions.
And because the engine underneath (agy) has its own toolbox, Gemini can pull its best tricks while it's in there — spinning up subagents to split a big design job, running design tooling, working several pieces at once. You hand off one task; it quietly runs a whole crew to finish it.
The trick: don't just delegate the work. Delegate the skills that make the work good, and let Gemini run them for you.
- Full command reference → SKILL.md
- Plain-English feature tour → docs/ask-gemini-overview.md
- Run the tests →
tests/run-tests.sh - License → LICENSE
ask-gemini turns one AI into a team — one for the eye, one for the hand — so your frontend gets the best design model's talent with none of the risk of letting it run loose.