# Lantern Keep

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [JIN WOO JEONG](https://cerebralvalley.ai/u/zzangbaba)
- **GitHub:** https://github.com/jwbaba-ai/lantern-keep
- **Demo video:** https://youtu.be/Eo5MEwkh_OE
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1216

Lantern Keep is a 1–4 player co-op castle defense game on Roblox. You defend the Lantern at the center of a keep while enemies push through four gates. Each match is one Day; win and the next Day gets longer and harder, up to a boss.

Two Meta models make the world react to the people playing it:

- AI director (Muse Spark 1.3). After every wave, the server builds a match summary: lantern health, which gate broke, each player's class, damage and deaths, revives, and the previous plan. During the 20-second intermission, Muse Spark returns the next wave as strict JSON (enemies per gate, tactic, event, a short director note). A validator checks enums, totals and day rules. The first wave of each new Day is designed from how the previous Day ended.
- NPC companions with memory (Muse Spark 1.3). After each match, the party's in-game events are stored, and Muse Spark turns them into memories and lines for the next match. Bram, the veteran captain, remembers the battle; Mira, the apprentice healer, remembers the people.
- Muse Image made the NPC portraits, the boss portrait and the event banners during development.

The game never waits on the model: any timeout, HTTP error, invalid JSON or rule violation plays a prebuilt fallback wave, and NPCs fall back to default lines. Every AI line is filtered and labeled as AI-generated, and only in-game events are stored.

Everything was built during the hackathon by one developer with Muse Code (implementation) and Claude Code (specs, review, Studio testing). Daily specs, test logs and asset origins are in the repository.

## More from Meta Global AI Developer Hackathon

- [Manage Assets Inc.](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1213)
- [WorldLab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1214)
- [Harold Cooper](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1215)
- [Pettigrew Lab](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1217)
- [Omnibasket](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1218)
- [LittleCreator](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1219)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1216. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt