Inspiration
I'm a product manager who can't really read code. Over the past six months I've vibe-coded and shipped 60+ small projects — games, tools, interactive toys. Around project 30, something started to bother me: I was shipping more and understanding less. Everyone talks about "outsourcing your thinking to AI." I was living proof.
Then I noticed the one thing every vibe coding session has: waiting. Your agent works for minutes at a time while you sit there scrolling your phone. Video games solved this decades ago — that's what loading-screen tips are for. So: what if your agent's work time became your learning time?
What it does
Osmosis is a learning companion that runs beside your coding agent. While Codex builds your project, Osmosis:
- Maps your project into a knowledge tree. Ask for a 3D webpage and it lays out what that actually means: Three.js → WebGL → scenes, cameras, meshes, lighting.
- Pops up lesson cards about the exact tech your agent is using right now — plain-language explanations written for non-programmers, each with a quick quiz, delivered in the dead minutes while the agent works.
- Lights up your skill tree as you answer, Duolingo-style.
- Remembers what you've mastered. Your profile follows you across projects — concepts you already know never pop up again.
You ship the project and keep the knowledge.
How we built it
[⚠️ PLACEHOLDER — rewrite with the real story after the build sprint, July 21 / 造完再写]
- Built end-to-end with Codex (GPT-5.6) in a single session — session ID submitted
- Node.js MCP server + SSE + a single-file web UI; GPT-5.6 generates the knowledge tree and every lesson card at runtime
Challenges we ran into
[⚠️ PLACEHOLDER —]
Accomplishments that we're proud of
[⚠️ PLACEHOLDER — ]
What we learned
[⚠️ PLACEHOLDER — ]
What's next for Osmosis
- Real spaced repetition — today, mastered concepts stop appearing; next, they return right before you'd forget them
- Cloud accounts so your mastery follows you across machines
- More coding agents — the MCP design is agent-agnostic
- Hardcore mode: your agent won't take the next task until you answer the current card. If you don't learn, your AI goes on strike.
