Built This Week/Builds

Builds

Mirror, Mirror on the Wall

A talking magic mirror is three steps in Google AI Studio: speech-to-text, a Gemini reply, text-to-speech. Ours worked on air, but it was laggy and named the wrong most valuable company.

For our Halloween episode, Episode 19 (October 31, 2025), Jordan built a talking magic mirror. You stand in front of the camera, ask "Mirror mirror on the wall…", and it answers out loud in a spooky voice. He built it in Google AI Studio with Gemini behind every step. He didn't say how long it took; Sam called it "a fun, quick build."

Why we built it

The week before, we had made some Halloween treats for Episode 18, including the AI costume builder. Since this episode dropped on Halloween itself, Jordan wanted to escalate. It was also a way to show off Google's new Build tool inside AI Studio, which we covered as news the week before and then used all week.

The stack

  • Google AI Studio (Build): generated and hosted the app from a prompt.
  • Gemini: answered the questions. Jordan at 04:08: "All of this is powered by Gemini."

How we built it, step by step

  1. Prompt the app in AI Studio Build. Gemini is already wired in, so there are no API keys or model plumbing to set up.
  2. Turn on the camera. The app asks for camera access so you see yourself "in" the mirror.
  3. Capture the question with real-time speech-to-text. On screen, your words appear as you say them.
  4. Send the text to Gemini for a reply.
  5. Speak the answer back with text-to-speech in a spooky voice.

How it turned out

It worked live. Jordan asked who was the smartest of them all, and then asked the mirror which company was the most valuable on the stock market that day. That week's news was Nvidia crossing a $5 trillion valuation, and the mirror missed it:

It said it was Microsoft here. It looks like it says that it's Microsoft instead of Nvidia.

— Jordan Metzner, Episode 19

He also admitted it was "a little slow and laggy." The pause between the question and the spoken answer was noticeable, which matters for something meant to feel like magic.

Sam's idea: Jordan lives on a street with a lot of Halloween traffic, so set up a monitor, webcam and microphone and let kids talk to the mirror. Jordan thought it was feasible with a monitor, a camera and some speakers. We did not report back on trying it.

What we'd do differently

  • Expect stale facts. The model answered from what it knew, not that day's market. Questions about "today" need live data.
  • Work on speed. The speech-to-text, model and voice round trip was the weak spot.

Later in the same episode, Jordan used AI Studio again to live-build the Guess My Costume app and the Trick-or-Treat route planner. To see how AI Studio stacks up against other builders, read Claude Code vs Google AI Studio.

FAQ

How do you build an AI magic mirror with Gemini?

Jordan built it in Google AI Studio's Build mode. The app uses the camera, turns your spoken question into text in real time, sends it to Gemini, and reads Gemini's answer back with text-to-speech in a spooky voice.

Is a Gemini talking mirror app fast enough to use live?

It worked on the show but was a little slow and laggy between the question and the spoken answer.

Can a talking mirror app be set up for trick-or-treaters?

Sam suggested putting it on a monitor with a webcam and microphone for kids on Halloween. Jordan thought a monitor, a camera and some speakers would be enough to pull it off, but we did not try it on the show.