Inspiration

Most of what we want to learn isn't a school subject. It's "make good cocktails for my party Saturday" or "don't get ripped off buying a used car." For that you get a generic course or a chatbot answer you forget by Tuesday. Duolingo makes learning stick, but only for languages. We wanted that for anything, built around your reason for learning. "Poko" comes from poco a poco, little by little.

What it does

Tell Poko what you want to learn. Poko, a companion who gets a new costume for every course, asks up to four quick questions and builds a personalized course:

  • A winding skill path of lessons, with optional side branches
  • Interactive slides built from nine components (diagrams, ratio bars, multiple choice, fill-in-the-blank, match-pairs and more)
  • Wrong answers explain themselves and let you retry. No lives, no penalties.
  • The first lesson is ready when you arrive, and the next ones are written ahead of time
  • Celebrations, a streak, and "Learn something different" to reshape a course without losing progress
  • Courses other learners built, ready to start

Every goal is safety-screened twice. If a goal is refused, Poko declines and nothing is stored.

How we built it

  • Expo + React Native + TypeScript client, installable on an iPhone home screen
  • Hono + Drizzle + SQLite server. It's the only place the API key lives.
  • Claude Opus 5.5 plans courses and writes lessons; Claude Haiku 5.5 screens goals. Every output is structured JSON, checked against a shared Zod contract before it's stored or shown.
  • Scoring and unlocking are plain code, never the model
  • Claude Code as our software factory: one shared session, sub-agents in separate worktrees, and GitHub Issues as the only record
  • Playwright runs full generated journeys and checks accessibility

Challenges we ran into

  • Structured-output limits: our slide schema was too big for the API's grammar, so we built a fallback to prompt-guided JSON. It still validates everything and retries once, feeding back the errors.
  • Wait time: we split planning from lesson writing, prefetch upcoming lessons, and use fast mode. Time to the first lesson went from about 60 s to 49 s.
  • A lesson could open before it was unlocked. Because lock state is computed on every read, never stored, the fix was in one place.
  • Playful, not loud: our first design was too serious and the second too neon.

Accomplishments that we're proud of

  • The full loop works live in any subject. Nothing is canned.
  • No generated HTML or code anywhere: model output is always data
  • Built in one day: 82 commits, 44 issues, 26 PRs, 143 passing tests
  • Lessons are reusable, so a completed lesson counts in every course that has it
  • No paid infrastructure beyond the API key

What we learned

  • Define contracts first: validated output is output you can build a UI on.
  • Store facts and derive state. That made reuse free and bugs easy to fix.
  • Let the model write and let code decide.
  • Agents need written context, and fewer tools beat more tools.
  • Time to the first lesson matters more than anything else.

What's next for Poko: Learn Anything #23

  • Native iOS and Android testing, then the app stores
  • Public hosting and accounts that sync across devices
  • Measuring real cost per course
  • More slide components and voice input

Built With

Share this project:

Updates

Submission history