The latest articles on DEV Community by QingQiu (@qingqiu). https://dev.to/qingqiu https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4122225%2Ff8b677c8-dcd3-4f0b-9a9d-3396a02d3325.png https://dev.to/qingqiu en QingQiu Sat, 12 Sep 2026 14:07:24 +0000 https://dev.to/qingqiu/i-built-a-choose-your-own-adventure-site-where-every-choice-is-an-ai-generated-video-2p92 https://dev.to/qingqiu/i-built-a-choose-your-own-adventure-site-where-every-choice-is-an-ai-generated-video-2p92 <p><strong>Live: <a href="https://playcyoa.com" rel="noopener noreferrer">https://playcyoa.com</a></strong></p> <p>A few weeks ago, AI video crossed a weird threshold: generation became <em>faster than playback</em>. Fal's H3 Max (a post-trained MiniMax H3) renders a 5-second clip in under 3 seconds, and Pieter Levels immediately turned it into Infinite Slop — an endless crowd-steered AI livestream that reportedly costs ~$4,000/day to run.</p> <p>I wanted the same "the video doesn't exist until you choose it" feeling, but for a single player — and at a cost a solo dev can actually afford. So I built PlayCYOA: choose-your-own-adventure stories where every branch plays a real AI-generated scene.</p> <h2> The trick: most players take the same branches </h2> <p>A 24/7 shared livestream burns money nonstop. A branching story doesn't have to: a depth-3 tree is only 7 clips, generated once (~$0.45 each at MetaSo's ¥0.09/sec pricing), then served from a static directory forever. Ten thousand people can play it for the same $3.</p> <p>The API is only called when someone writes their <em>own</em> premise — and that's where the cost guardrails live:</p> <ul> <li> <strong>A hard daily generation cap</strong> (currently 30 scenes/day, one env var)</li> <li> <strong>Prepaid provider balance</strong> — when it's gone, generation just pauses; the site and all cached stories keep working</li> <li>If traffic ever explodes, the expensive action is already the gated action</li> </ul> <h2> Stack </h2> <ul> <li> <strong>Next.js 16</strong> — static genre/story pages for SEO, API routes for the generation proxy</li> <li> <strong>MiniMax H3 via MetaSo</strong> — MiniMax-compatible API at roughly 1/6 the official price; the provider is abstracted so I can swap to fal or MiniMax direct with one env var</li> <li> <strong>An LLM</strong> for writing the two branch options in custom mode (with static fallbacks when no key is set)</li> <li>A small Python script that expands a premise into a branching tree (LLM) and renders every node's clip offline — content production is a batch job, not a runtime dependency</li> </ul> <h2> What I learned </h2> <ul> <li> <strong>Async branching is the sweet spot.</strong> Real-time infinite streams are a rich team's game right now; letting a single user wait ~40s per scene works today and the UX holds up.</li> <li> <strong>Continuity is the hard part.</strong> Each clip is generated independently, so characters drift between branches. Framing each scene as a <em>location change</em> ("follow the candlelight upstairs") hides most of it — the cut becomes a feature.</li> <li> <strong>Video SEO pages are thin on the ground.</strong> The SERP for "AI choose your own adventure" is mostly text-only tools and tiny new sites — felt like an open lane.</li> </ul> <p>If you try it, I'd love to hear where the story continuity breaks for you — that's the thing I'm iterating on next.</p> showdev ai webdev nextjs