hestermani

Letra Letra is a 3D letter-learning adventure for pre-K kids ages 3–6. Players walk through a googly-eyed alphabet world, bumping into letter characters to hear them, spelling missing pets, and matching phonemes to letters. Three modes ship: Spell the Word, Find the Alphabet, and Match the Sound, plus a trophy shelf for repeat play. Before each game, the kid picks an avatar, a biome, and the letter case: Avatars: Kid, Car, Rocket Biomes: Park, Moon Letter case: UPPERCASE, lowercase, Mixed Plays in any browser, installs as a PWA, and works offline. Try it: https://www.playletra.com The Problem Most pre-K alphabet apps are flat 2D tapping games with stock TTS, and most ask non-readers to read on-screen instructions just to navigate. Letra is the opposite: a tactile 3D world where every prompt is voiced, so a child who cannot read yet can still drive every screen alone. Big buttons, generous proximity collection, voice-on-hover menus, and a hint timer that gently re-orients instead of scolding, all designed around one core assumption: the user cannot read yet. ElevenLabs ElevenLabs voices every line in the game: every letter name, every phoneme, every prompt, and every celebration, all in one warm, consistent custom narrator voice, Marissa. The full corpus of roughly 90 short MP3s is pre-baked at build time, so a child’s 40th replay costs zero tokens and starts instantly from cache. Token economics flip when the audience replays, and a 3-year-old will replay the same letter 40 times in a session. Without ElevenLabs, the game falls back to the browser’s Web Speech. It works, but it does not feel like a friend. Zed Zed’s agentic editor wrote the majority of the React shell and the three.js engine, including geometry, lighting, and the per-frame loop. It also handled the input layer, including keyboard, gamepad, and on-screen joystick support, plus the trophy state machine and the biome registry. I drove the design and kid-UX judgment calls, (with help from my four-year-old assistant): proximity collection radius, hint phrasing, joystick ergonomics, and voice-on-hover. ⸻ Source: https://github.com/billums123/letra

Submitted 30 Apr 2026

STADIUM turns any walk, run, or bike ride into a live AI sports broadcast. Tap GO and two ElevenLabs voices (a play-by-play announcer and a color commentator) narrate your pace, distance, and goal progress in real time, layered over an ElevenLabs-generated stadium crowd and a cinematic music bed that swells for the final dash. The problem is simple: walking and low-intensity training are how most people are supposed to build their health, but they feel unrewarding, boring, and invisible. You grind without feedback. You quit. STADIUM treats every step like it matters, because for the person moving, it does. The app reacts to real motion: pace surges trigger excited commentary, goal progress drives urgency, and the final dash plus victory horn make crossing a finish line feel earned. It is psychology as a product , turning intrinsic effort into externally validated moments. The bigger pattern is that this same loop works far beyond walking. It can apply to physical therapy recovery sessions, rehab walking programs, strength-training sets, interval running, cycling goals, meditation streaks, and any self-improvement domain where progress is real but feedback is delayed or invisible. A personal live broadcast fills the motivation gap that generic fitness apps leave behind. I picked walking first because it has the lowest barrier to entry and can yield extremely positive health benefits. ElevenLabs powers every audible surface: • TTS (eleven_turbo_v2_5) for the two voices • Sound Generation for the crowd loop • Music for the cinematic bed • Voices API for a voice picker that lets users choose their own announcers That is four ElevenLabs products, each doing load-bearing work. None of it is decoration. Kiro drove the workflow. Every feature started as a spec — /.kiro/specs/stadium/{requirements,design,tasks}.md — before a line of code was written. Steering docs in .kiro/steering/ kept the agent consistent across sessions, and agent hooks enforced that every commit traced back to a requirement.