200 points · 1 submission
with Zed
Soundle is a mobile-first word-guessing game where every clue is audio — no images, no text hints, just sound. Players listen to 2–3 audio clues and build the hidden word by tapping letters from a scrambled tile bank. Clues range from raw sound effects ("a wall clock ticking") to AI-narrated descriptions ("a flat panel of buttons you press to type") — combined, they reveal the answer: CLOCK, KEYBOARD, RAINBOW, and 17 more across 4 themed packs. The problem it solves: Word games are visually dominated. Soundle makes audio the primary interface — training your ears instead of your eyes, and creating a genuinely novel puzzle format. How it uses ElevenLabs: Every audio clue is generated by ElevenLabs using two APIs: - Sound Generation (`/v1/sound-generation`) — produces natural sound effects and music background from text prompts like *"friendly dog barking twice, clean studio audio"*, with a custom prompt wrapper that explicitly strips out any accidental speech. - Text-to-Speech (`/v1/text-to-speech`) — narrates description and word-part clues using the `eleven_multilingual_v2`. How it uses Zed: The entire project was built in Zed. Zed's AI agent was used throughout — most critically to design and wire up both ElevenLabs integrations inside a single Next.js API route, generate the batch audio script, and iterate on the puzzle data schema. Zed's speed and inline AI made it possible to go from zero to a working audio-driven game in a single hackathon sprint.
Submitted 30 Apr 2026