iTranslated by AI

The content below is an AI-generated translation. This is an experimental feature, and may contain errors. View original article
🛸

[Antigravity] Building a Coffee Chat Topic Generator from a Single Doodle

に公開

I recently decided to host an online chat session at work, so I built a web app where an AI generates casual conversation topics to help keep things flowing. 🪄◝*ෆ

Gemini generates chat topics for me like this

I decided to let the trending "Antigravity" handle the development this time, and including minor tweaks, I was able to get it deployed in about an hour and a half ⭑.ᐟ

About Antigravity

It is an IDE released by Google in November 2025:

https://antigravity.google/

I don't know all the technical details, but it describes itself as an "AI-agent-led coding environment," allowing you to code while leveraging models like Gemini and Claude Sonnet ⸝⋆

༝༚༝༚…🖋️ These articles provided a clear overview:

https://ramble.impl.co.jp/11795/
https://ramble.impl.co.jp/11853/

About what I built

It is a tool where Gemini (Gemini 2.5 Flash) generates chat topics based on the following rules 🧸💭

prompt:
You are an assistant tasked with livening up an online casual chat session (once a week, 30 minutes, IT company, judgment-free open space).
Please provide exactly one "casual chat topic" that is fun and encourages sharing thoughts, allowing participants to relax while holding a cup of coffee.
Return only the text of the topic (e.g., in a question format like "How about...?"), and no greetings are necessary.

- When "Give me a random topic!" is selected: Think of a completely random and unexpectedly interesting topic.
- When a category is selected: Think of a topic related specifically to the "xxx" theme that is easy for people in the IT industry to answer.
- When a topic is specified by the user: A keyword "xxx" has been provided by a participant. Based on this keyword, think of a positive topic that is easy for everyone to join in.

༝༚༝༚…🖋️ I had the details on the technologies used written out in the README, so please refer to that

https://github.com/raynya1212/coffee-chat-generator


The initial doodle

When you're ready to do some "vibe coding," you might want to give instructions like:
🦥< Hey, build me xxx

But this time, I started by sketching out an image of what I wanted the tool to look like⋆ 𖦆 💭

I just drew it roughly in Windows Paint, thinking "something like this would be nice."
I originally imagined a cat holding coffee, but I changed it later

The point here

At this stage, having at least a rough idea of:

  • What I want to do (I want to make a tool like this⭑.ᐟ)
  • User flow
  • Design concept

made the subsequent process feel much smoother ꒰*ੈ🌙

Of course, AI can build things with just a simple prompt like:

🦥< Hey, build me xxx

But it often ends up being "not quite what I had in mind..." Correcting from that "different" state is quite difficult, and repeated deployments can often lead to errors... so I felt it was important to solidify the image of what I really wanted to achieve ⸝⸝ ˚. 🌙

Expanding on the UI image

I had a general idea of the UI, but I used Stitch to clean up the doodle above and flesh out the concept:
https://stitch.withgoogle.com/

This is also a tool released by Google (technically acquired by them), specialized in AI-generated UI design

https://x.com/arnaudai/status/1924942577545195982

I fed my doodle into Stitch and explained the kind of web app I wanted to build, and it came up with this design ५✍️

Oh, how cute

Developing with Antigravity

Now that I had a rough idea of the image, it was time for the actual development ┊꒷⛏️⑅

1. The Prompt

I provided the following as an initial prompt 🪄︎︎◝✩
*I ran this in "Planning" mode.

I want to build a web app that generates conversation topics for online casual chat sessions.

Assumptions for the online chat session:
- Once a week, 30-minute sessions
- Connect online
- It is a judgment-free open space where we can talk
- Since it is an IT company, they might enjoy IT-related topics

What I have in mind:
- Use Gemini AI (I have the API key ready)
- For the top page UI, refer to the image at the path below:
xxxxx/xxxx.png

- User flow:
1. The user selects either [Think from scratch] or [Choose from topics].
2. If [Think from scratch] is selected, the AI generates a chat topic from scratch.
3. [Choose from topics] allows the user to select from themes like "Movies," "Weekend outings," "Gourmet," "IT news," etc., and the AI generates a topic for conversation based on that choice.

- UI suggestions:
・I want the menu items like [Think from scratch] to have a playful English tone (something like "give me a xxx").
・Since it's a "coffee chat" theme, a coffee motif would be nice.
・If there is a loading state, using a Lottie animation would be good (if we could animate the cat itself, that would be great!).

2. Refining the Plan

Once Gemini finishes thinking, it creates an Implementation Plan 🧾.

You can review what the construction looks like here and add comments for anything you want to address ५✍🏻

When you click the [Review] button, Gemini checks your comments and updates the plan. You repeat this a few times, and when you press the [Proceed] button, it advances the implementation to a state where it runs locally 👟ˊ˗

3. Making Tweaks Until Satisfied

I checked the behavior locally, and for the parts I didn't like, I added instructions in the chat ✒️ᝰꪑ

However, this time I only had to add/change a few things:

  • Animation settings (Initially it was an animation of emojis floating around, but I used a cute Lottie animation instead)
  • Font settings (Specified Google Fonts)
  • Category settings (Adjusted after actually seeing them)
  • Added Basic Authentication to the page
  • Handling for when the Gemini rate limit is reached (Decided to return pre-prepared themes)

Since I am just giving instructions in words for these changes, I didn't have to rewrite code by hand ๋⭑🛸๋⭑

4. Deploying to Vercel

I was satisfied with the behavior, so I deployed it using Vercel and ta-da! It is complete 🐾✧˖°.

Thoughts

From a vague doodle to a working web app, it happened in the blink of an eye!
Actually, writing this article probably took more time... 🤧ˎˊ˗

Compared to GitHub Copilot

Regarding other AI-powered IDEs I have used, there is GitHub Copilot which runs inside VS Code:
https://github.com/features/copilot?locale=ja
Compared to that, I felt that Antigravity's ability to show:

  • Task
  • Implementation Plan
  • Walkthrough

was quite impressive *⋆✈️💭 Since it clearly displays the flow of progress and what the implementation will look like, it is easy to avoid getting lost.

༝༚༝༚…🖋️ These appear to be what are called Artifacts generated in Planning mode:
https://antigravity.google/docs/artifacts

Also, since Antigravity is a fork of VS Code, the workflow of the editor and chat with AI is mostly the same.

Conclusion

I hope this was a little helpful 🪽₊˚.ꪆ.

Discussion