Karpathy — LLM Wiki¶
- URL / path: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
- Consulted on: 2026-04-17
- Commit / version: gist as retrieved on 2026-04-17
- Status: active
Why we read it¶
Decide the shape of docs/wiki/. We already had a single
lessons-learned.md file; we needed a pattern that scales as the
repo accumulates design notes, decision history, and material
extracted from vendor repos (vendor/kimi-cli, vendor/bub) and
external references.
What we took¶
The three-layer architecture verbatim:
- Raw sources (immutable) — for yaya this maps to
vendor/kimi-cli,vendor/bub, external gists and articles, the homework PDF. Agents read; they never modify. - The wiki —
docs/wiki/— LLM-authored, maintained markdown pages that compile knowledge extracted from sources. - The schema —
docs/wiki/AGENT.md— tells agents how to maintain the wiki.
The three operations (ingest / query / lint), the two special
files (index.md, log.md), and the log-entry prefix convention
(## [YYYY-MM-DD] <kind> | <title>) — all adopted.
Page shapes (source / concept / decision / lesson) in
docs/wiki/AGENT.md are yaya-specific instantiations of the "pick
what's useful" clause at the end of the gist.
What we rejected¶
- Obsidian / Marp / Dataview tooling. yaya's audience is engineers reading on GitHub or in their editor; Obsidian-specific extensions would fragment the experience.
- Browser extensions (Web Clipper) and image handling. Our sources are code repos, text docs, and specs — no images to download, no web clipper needed.
- Dedicated search engine (qmd). At current scale
index.mdis enough. Revisit when the wiki crosses ~100 pages. - Dataview / YAML frontmatter metadata. Premature. Markdown + relative links suffices until we have a concrete query the plain text can't answer.