FrogQuest: I turned my to-do list into an 8-bit RPG where I'm the hero
▶️ Try the Space · 🎬 Watch the demo
I kept putting off my hardest tasks, and the productivity classic Eat That Frog says to do the opposite: face the biggest, ugliest task — the frog — first. So I built FrogQuest: tell it your day, your goals, your chores, and it forges them into a text-based RPG quest log where every task is an 8-bit quest, your hardest task is the boss fight, and the hero in every scene is you — it puts you in the art from a photo.
Finish a task, and your quest image transforms into a victory scene. Couldn't do it? Tell the Frog Master why, and the scene becomes a forgiving retreat — the hero lives to fight another day. No guilt mechanics. Productivity apps that shame you don't get opened twice.
How it works
The architecture rule that made this shippable: the LLM never writes code, markup, or pixels — it only writes JSON.
- NVIDIA Nemotron 3 Nano 4B (GGUF, via llama.cpp) is the quest designer. It reads your goals, applies Eat That Frog — picks the frog, chains big goals into quest lines, sprinkles in optional self-care bonus quests — and emits a strict JSON quest log (schema-constrained output, validated and clamped before anything renders).
- FLUX.2 [klein] 4B (Black Forest Labs) makes the art. Each quest gets an initial scene generated with your photo as a reference, so you're the hero. Success and failure don't regenerate the image — they edit the existing one, so it's the same hero, same scene, new outcome. That's what makes it feel like a story instead of a slideshow.
- A fixed renderer turns the JSON into the retro UI. Deterministic. Nothing the model can break.
Both models run on the Space itself via ZeroGPU — no external APIs. Total model weight: ~8B parameters across two 4B models.
The feature I'm most proud of: quest planning with web research
If you have a goal but no idea where to start — "learn embedded programming," "run a 10K" — turn on web research in your campaign settings. The AI researches the topic and plans the quest chain for you, breaking the fuzzy goal into concrete, ordered quests. You stop staring at a mountain and start at step one.
What I learned the hard way
FrogQuest is a pivot — I wrote up the failed first attempt here. The first version tried to make a small model generate playable games from scratch. It looped, it overflowed its context, and when it did produce games, they barely worked — it used the 30B version of the same Nemotron model writing, but correct game logic in html in one shot is just past its ceiling, and no amount of prompting fixes a capability limit.
The lesson that unlocked everything: stop asking a small model to do the hard thing; redesign so the hard thing doesn't exist. Language models are great at language — narrative, theming, structure. So the model writes quests and image instructions as data, and tested code does all the rendering. Small models aren't weak — they're specialists. Build around what they're genuinely good at and they shine.
Stack
Gradio (custom retro UI) · ZeroGPU · llama.cpp (Nemotron 3 Nano 4B, Q8) · diffusers (FLUX.2 klein) · browser-only storage — your photo and quests live in your browser, not my server.
Try it
Forge your to-do list into a campaign: FrogQuest on Spaces. If it makes your Monday list look less terrifying, leave a ❤️ on the Space and I'd love feedback in the comments either way. 🐸
