Apr – Jun 2026 · age 17 · web app (paused)
Interacta
A social feed where every post is something you can play: swipe up, and the next mini-game, puzzle or learning tool fills the screen. An idea I carried for two years, built in six weeks.
- Status
- Paused, live demo works
- Pace
- 400+ commits, ~73 PRs in 6 weeks
- Platform
- Installable PWA, phone-first
Where the idea came from
I had this idea about two years before I built it. Back then I spent a lot of nights imagining I was running Microsoft, deciding what it would take to get back into the race with the big platforms. Social was the side where it had nothing. Interacta was my answer for that side: short-video feeds are great at holding attention and give nothing back, so what if every card was something you play, solve or learn from instead of just watch?
yes, a teenager pretending to be a CEO at 2 AM. it's how most of my ideas start.What I built
- A vertical, full-screen swipe feed of playable widgets: mini-games, puzzles and learning tools.
- Anyone can post a widget: write the code, start from a template, or describe it and let AI build it.
- Remixes, personal bests, challenges between friends, simple multiplayer rooms.
- Accounts, follows, direct messages, notifications, moderation and an admin dashboard.
- An installable progressive web app, phone first.
How I worked with AI
Honest version: I didn't type most of this code. AI agents did, and anyone can ask an AI for code. What I did was run them like a small team. I decided what the app had to do and how it had to feel, broke it into tasks, and checked every change before the next one started.
The models I had in spring 2026 forgot everything between chats and drifted when a task got long. So I built a system around them:
- Rules in files, not in chat. Project rules, a widget build prompt and safety notes lived in markdown files that every new session read first.
- Nothing I asked for got lost. I mined all 74 chat sessions into one "directives" file, so an instruction given once was never forgotten in the next chat.
- Small, reversible steps. One task per commit, work through pull requests (about 73 in six weeks), so any mistake could be undone on its own.
- A gate for AI mistakes. Lint, tests and a build in CI had to pass before anything merged, to catch the undefined names and crashes AI code tends to leave behind.
The first prototype was made with Lovable, an AI app builder. After that, the product decisions, the rules and the testing were mine; the typing was mostly the AI's.
The parts that are mine
How it feels
- No spinning circles. Loading shows skeletons shaped exactly like the content that's coming.
- Loaders wait 1.5 seconds before appearing, so on a fast network nothing flashes at all.
- Every part of the app loads on its own, from cache when it can, so what's ready shows up right away.
- Messages appear instantly as "sending", then turn solid when they're sent. The keyboard stays up for the next one.
- No start screens: the first frame of a widget is its thumbnail, and you're already playing. Control hints fade in and out.
- Errors in plain language that say what to do next, never an error code.
Safety, when anyone (or any AI) can post code
- Every widget runs in a sandboxed frame that can't reach the app, the account or other widgets.
- A scanner checks code before it's published, blocks anything trying to escape the sandbox, and explains why in plain words.
- A 3-second test run on publish: at least 20 FPS and zero errors, or it doesn't go live.
- A watchdog freezes a widget that stops responding, and a crashing widget can't take the feed down with it.
- The AI builder bakes these safety and size rules into every prompt it writes, so AI-made widgets pass the same checks.
- Database rules so people can only read and change their own private data, even by editing a link. Rate limits, reports that auto-hide a widget, and daily backups.
What I learned
People open social media to switch their brain off. A feed that asks you to play something asks for effort, and most people scrolling don't want effort. So a feed like this only works if the content is really good, and the only way to be sure of that is to make the content yourself, or with a few trusted makers, and let people just scroll. I built the platform first and hoped the content would come. Next time it's the other way around.
Scope. A social network is several products at once: a feed, accounts, content, moderation. I got the core working, then paused for my final school year before every part was finished.
paused, not abandoned. probably.Stack
- React
- TypeScript / JavaScript
- Tailwind CSS
- Supabase
- Postgres · PL/pgSQL
- PWA
Links
Development is paused. The live demo works, but some features are unfinished. Best on a phone. The source is private.