A skill is documentation that learned to do things.
| K-Line | Why Related |
|---|---|
| moollm/ | This IS the core of MOOLLM |
| prototype/ | Skills are factories (Self language) |
| incarnation/ | Gold-standard skill instantiation |
| play-learn-lift/ | LIFT produces skills |
| sister-script/ | Document → Procedure → Automation |
| sniffable-python/ | LLMs understand from first 50 lines |
| plain-text/ | Everything persists in files |
| speed-of-light/ | Many agents, many turns, one LLM call |
| constructionism/ | Build to learn (Papert) |
| empathic-templates/ | Smart generation, not string substitution |
Quick Links:
- Full Specification — complete protocol
- Interface Card — machine-readable methods
- Delegation Protocol — Self-like inheritance
- Instantiation Protocol — how skills become instances
The meta-protocol: how skills work, evolve, compose, and how MOOLLM advances the state of the art.
A skill is a documented capability that can be:
- Instantiated — Create running instances with state
- Composed — Build complex from simple
- Inherited — Clone and override (Self-style)
- Lifted — Extract patterns into reusable skills
| # | Innovation | What It Means | Proof |
|---|---|---|---|
| 1 | Instantiation | Skills as prototypes that create instances | adventure-4 from adventure/ |
| 2 | Multi-Tier Persistence | Platform → Narrative → State → MOO-Maps | Session logs + world state |
| 3 | K-lines | Names activate entire knowledge contexts | "Palm" → all of Palm's soul |
| 4 | Empathic Templates | Smart generation, not string substitution | Character descriptions |
| 5 | Speed of Light | Many agents, many turns, one call | 33-turn Fluxx, 21-turn cat prowl |
| 6 | CARD.yml | Machine-readable interface with advertisements | Every skill has sniffable metadata |
| 7 | Ethical Framing | Room-based inheritance of performance context | Pub framing → all children inherit |
skills/
my-skill/
README.md # Human entry point (GitHub renders)
SKILL.md # Full spec with YAML frontmatter
CARD.yml # Machine-readable interface
*.tmpl # Templates (optional)
Every skill has:
- README.md — Human discovery (you're reading one!)
- SKILL.md — Full protocol with YAML frontmatter
- CARD.yml — Interface: methods, tools, state, advertisements
Documentation → Procedure → Script → Tool
↑ ↑ ↑ ↑
PLAY LEARN LIFT SISTER-SCRIPT
Play-Learn-Lift: Do it manually → Notice patterns → Extract skill
| Tier | Where | Lifespan |
|---|---|---|
| Platform | Cursor session | Ephemeral |
| Narrative | LOG.md data islands |
Read-mostly |
| State | *.yml files |
Read-write |
Key patterns:
- Data islands: Embed YAML in logs with
#object-idaddressing - Promotion: Pop to
.ymlfile when editing needed - Log inheritance:
inherits: LOG.md#birth-state
LLMs can understand skills by reading first ~50 lines:
- Lines 1-15: YAML frontmatter (name, tier, tools)
- Lines 16-25: Purpose and philosophy
- Lines 26-40: File map (what's in the skill)
Efficient discovery without reading everything.
| Demonstration | Turns | Agents | What It Proves |
|---|---|---|---|
| Stoner Fluxx | 33 | 8+ | Complex game state, rules, consistency |
| Cat Prowl | 21 | 10 | Parallel agents, coordinated behavior |
| Palm Incarnation | 1 | 6+ | Committee debate, autonomous creation |
The architecture works. Speed-of-light simulation is real.
"Start with jazz, end with standards. But never stop playing."