# .wiki/ — Karpathy LLM Wiki

Curated knowledge base. Structure:

- `entities/<slug>.md` — physical/logical assets (hosts, NAS, VDS, router, recovery-VM)
- `concepts/<slug>.md` — patterns, root-cause analyses, recipes, postmortems
- `packages/<slug>.md` — third-party / first-party software packages relevant to this stack
- `sources/<slug>.md` — chronologies, session traces, primary records
- `raw/` — immutable external clippings (transcripts, screenshots, vendor docs)
- `index.md` — catalog of all pages
- `log.md` — op-log of ingest / promote / lint operations
- `overview.md` — high-level project orientation

## Domain conventions

### Mandatory: read design context before implementation

Before picking up any task in `.tasks/`, load the full design context. It does **not** live in this repo as a standalone source — only pointers do. Sources, in order:

1. **Local design (canonical):** `.wiki/concepts/admin-infra-project.md` — ingested via promote 2026-05-21. Identity, scope, content inventory, migration recipe (subtree-split + read-tree merge for populated prefixes), initial agenda.
2. **Brainstorm process trace (rationale):** `~/projects/.workshop/.archive/2026-05-21-admin-infra-project.md`. Why each decision was made (recommend vs menu trade-offs), what was rejected (filter-repo vs subtree-split, "admin-only" skills, splitting roadmap from ops), anti-patterns flagged during brainstorm.
3. **Local `overview.md`** — thin summary, quick orientation only — never the source of truth.

Do **not** invent migration recipes, file lists, taxonomies, or scope decisions from task `where_stopped` lines alone — those are pointers, not specifications. The concept doc and archive contain the rationale.