Stop bulk-loading SKILL.md into your agent. Make AI knowledge addressable, retrievable, and semantically typed — on demand.
Three architectural inversions
Agents see a ~3 KB index of what exists. Atom bodies load only when the brief actually calls for them — addressed by ID, not guessed by relevance.
14 edge verbs (requires, validates-with, contradicts, specializes…) carry semantics the parser and L3 checker reason over. The graph is not decoration.
Each atom compiles to summary (~30 tok), core (~150 tok), full (~380 tok). Pick the smallest projection that fits the budget.
The atom format
Every unit of knowledge is addressable by ID and declared in an edge graph. The compiler understands what it knows, what it depends on, and what contradicts it. Agents retrieve only what the brief needs.
fact WaterBoilsAt100C {
id: "@example/fact-water-boils-at-100c"
version: "1.0.0"
statement: "Pure water boils at 100°C at 1 atm."
confidence: 0.99
domain: physics
related: [
@example/term-celsius,
@example/rule-altitude-affects-boiling,
]
validates-with: [
@example/source-nist-water-properties,
]
}The agent sees what exists. It retrieves what the brief demands. Nothing more, nothing less.
PRIME-PROTOCOL v1.0 frozen as of 2026-05-07. MCP server ships for Claude Code, Cursor, and any MCP-capable client.