A coordinate-addressed, navigable, self-maintaining tree over a flat skill directory: tree/map/search plus the discover/cohere/emit/unemit coherence loop. The working backing for Paper 1's correction (v0.2.0, 62 tests passing).
An executable tree surface for a language model—typed nodes navigated by descent, runnable at the leaves, exposable over MCP. Noted for the dated precedence only: it was public in August 2025, roughly two months before agent skills shipped and before on-demand loading was advocated over upfront schemas.
The graph-native form in minimal, runnable shape: skills as hypergraph nodes; typed edges (specializes/delegates_to/must_follow/contains) as state transitions that inject context; chart() derives tree-projections on demand. A runnable existence proof, not a production substrate.
Write a reasoning persona you can actually check. One command, zero deps: define a persona as an ordered list of moves with cue-phrases, then gauge any transcript against it—did the prose perform the moves in order, or scramble them.
Turn a one-line attention-chain into a reusable attention-directing block. Compiles [problem] => [constraints] => [options] => |decision| into the prompt block you actually inject.
Bundle skills + a persona into a droppable agent—one command. A folder of skills and a one-line persona become a self-contained, installable package with the skills mapped and coordinate-addressable.
Coordinate-addressed mapping for skill folders; released via the research monorepo; standalone release forthcoming.
Coordinate-scoped BM25 search over a coordinate-mapped skill tree. --scope 0.2 ranks only within that branch and below: the coordinate is the address and the search region.
Recover the human steering flow from a conversation arc: extract the sequence of human inputs from session transcripts, clean it, and turn it into a reusable move-pattern plus a flow-vs-aside map.
Zero-config full-text search over any folder. One command. No vector DB, no model, no dependencies—SQLite FTS5 / BM25, stdlib only.