MOOLLM -- treating the LLM as eval() for a microworld operating system.
Started this incarnation on Dec 30, 2025 -- but it's the crystallization of decades of earlier prototypes, all the way back to my Commodore-64 Logo Adventure. Built on top of Anthropic's Skills framework, extended with seven innovations (and counting):
1. Instantiation -- Skills as prototypes creating instances with their own state
2. K-lines -- Names as semantic activation vectors (Minsky's Society of Mind)
3. Empathic Templates -- Smart generation based on semantic understanding, not string substitution
4. Three-Tier Persistence -- Platform (ephemeral) → Narrative (append) → State (edit)
5. Speed of Light -- Many turns in one call, minimal tokenization overhead
6. CARD.yml -- Machine-readable skill interfaces with advertisements
7. Ethical Framing -- Room-based inheritance of performance context
Lineage: Colossal Cave → TinyMUD → LambdaMOO (filesystem as world). Papert's Logo and constructionism (learnable microworlds). Will Wright's SimCity and The Sims (I worked on the originals) -- the "Simulator Effect" where players imagine more than you simulate, and the SimAntics visual behavior programming language.
YAML Jazz: Comments aren't ignored -- they're semantic. The LLM reads and interprets them. A comment like "# gentle but firm" on a character trait actually affects behavior. This inverts the traditional "comments are for humans" assumption. Comments become part of the program and data.
The core idea: instead of prompt engineering, you give the LLM a github repo filesystem to inhabit: a persistent microworld. Seymour Papert's Constructionist philosophy comes alive, with Minsky's K-Lines pulling the strings. Skills are programs (not documentation). Characters have persistent state in directories, and can reflect on and edit themselves. Everything is inspectable and editable by human AND model. Model and platform independent. Runs on Cursor and other tools and orchestrators.
The proof is in adventure-4 -- a complete text adventure with 150+ files, 6000+ lines of session transcripts.
Repo: https://github.com/SimHacker/moollm
MOOLLM Manifesto: https://github.com/SimHacker/moollm/blob/main/designs/MOOLLM...
The MOOLLM Eval Incarnate Framework: https://github.com/SimHacker/moollm/blob/main/designs/MOOLLM...
Adventure 4 Example: https://github.com/SimHacker/moollm/tree/main/examples/adven...
My sessions as proof it works: https://github.com/SimHacker/moollm/tree/main/examples/adven...
79 Anthropic Skills (standards compatible, plus extensions, intertwingled with k-lines) and growing: https://github.com/SimHacker/moollm/tree/main/skills
A guided tour through the MOOLLM skills and microworld -- Session Log: K-Line Connections Safari: https://github.com/SimHacker/moollm/blob/main/examples/adven...
Adventure Compiler Design Discussion -- Adventure Uplift Session Log: https://github.com/SimHacker/moollm/blob/main/examples/adven...
MOOLLM Kernel: https://github.com/SimHacker/moollm/tree/main/kernel
Happy to answer questions about any of the weird design decisions!