# Jermaine Dennis

- **Event:** [AIE Code Summit 2026 Hackathon](https://cerebralvalley.ai/e/aie-code-2026-hackathon)
- **When:** Sat, Nov 7 at 9:00 AM – 10:00 PM (PST)
- **Where:** San Francisco, CA
- **Team:** [Jermaine Dennis](https://cerebralvalley.ai/u/jdshaolinstar)
- **GitHub:** https://github.com/jdshaolinstar/rsi-game
- **Demo video:** https://youtu.be/GH36qRGVspc
- **Gallery:** https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/64

Live demo: http://patchwork-demo.us-east-1.elasticbeanstalk.com (control room, view-only: /control)

Patchwork is a game that improves itself. A playable dungeon ships with bugs, and a coding agent on the Meta Model API keeps making it better while people play. It also learns how to code better from its own mistakes.

Players' crashes and feedback, and goals the creator types in plain words, go into one inbox. Muse Spark routes each item: gameplay and bugs go to the code Fixer (muse-wonder-1.0), and "it's too grey" goes to a vision loop. Every change has to earn its release:
- A code patch must improve the goal score on held-out bot playtests it never saw.
- It must pass guardrails. A patch that deletes a room to "help more players reach the boss" doubled the score and was still rejected.
- It then ships as a canary to half the players and rolls back if they crash more.
- A new look ships only if Meta's vision model prefers it over the old one in both image orders, toward target art made with muse-image-1.0.

Every version is self-contained and playable from a picker, with a "what changed" note listing the player reports it fixed.

The agent also improves itself. Each failed patch is a lesson. A Reflector writes a new skill as a versioned file (`skills/skill_NNN.md`). A new skill manifest is adopted only if it fixes more bugs on a held-out bench the Reflector never sees. In our logged run, the Fixer went from 0 of 6 held-out bugs to 6 of 6 after it wrote and adopted the skill "don't stop at the reported symptom, trace the whole path".

NEW GAME FROM RSI then builds a brand-new game twice, with the learned skills and best version and without them, and playtests both. The whole system can be saved and reloaded as checkpoints stored in S3.

Built today with Claude Code. Meta models: muse-wonder-1.0 (patches, skills, new games), muse-spark-1.3-contributor (vision judging, routing, goal mapping), muse-image-1.0 (target art).

## More from AIE Code Summit 2026 Hackathon

- [LAZARUS-TWIN](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/61)
- [Sift](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/62)
- [darkmatter: ebert-spielberg-lucas-bohr](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/63)
- [Lazarus](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/65)
- [featherAI](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/66)
- [Receipts](https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/67)

---

Markdown version of https://cerebralvalley.ai/e/aie-code-2026-hackathon/hackathon/gallery/64. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt