I wanted an entry for the Hacktoberfest open-source AI challenge that actually gets people off the couch. Each day the app deals five quests (2 easy, 2 medium, 1 hard) plus one shared quest everyone gets on the same day. You open the live camera, take a photo, and CLIP compares it against the target plus a set of negatives. A 3-minute cooldown after each check stops couch farming.

Tech: Single-file vanilla HTML/CSS/JS PWA (`index.html` + `styles.css` + `sw.js`, no backend, no account). Photo check is `Xenova/clip-vit-base-patch32` zero-shot classification, quest text is Gemma 3 270M falling back to Qwen2.5 0.5B then templates — all via transformers.js from jsDelivr, model IDs as constants at the top of the script. XP is 10/20/40 by difficulty plus 50 for the shared quest, streak multiplier x1.0–x2.0, away bonus up to +20. Plant grows at 0/100/400/1000/2500 XP and droops after 3 days away. Progress in `localStorage` under `tgq`, service worker caches the app for offline use after first load.

What I learned: Small on-device models need guardrails — the quest writer drifts off-topic so its output is validated with a target-mention check and falls back to templates. Framing matters a lot for CLIP: same tree too far or too close flips the result. Honor-system limits are real too — you can still photograph a screen, and real anti-cheat would need a server.

Next: Better framing hints before the snap, per-quest confidence tuning, and a shareable Grass Dex card.