Karpathy LLM Wiki

Karpathy LLM Wiki

Build and maintain a personal knowledge base using LLMs. You manage two directories: raw/ (immutable source material) and wiki/ (compiled knowledge articles). Sources go into raw/, you compile them into wiki articles, and the wiki compounds over time.

Core ideas from Karpathy:

• "The LLM writes and maintains the wiki; the human reads and asks questions."
• "The wiki is a persistent, compounding artifact."

Architecture

Three layers, all under the user's project root:

raw/ — Immutable source material. You read, never modify. Organized by topic subdirectories (e.g., raw/machine-learning/).

wiki/ — Compiled knowledge articles. You have full ownership. Organized by topic subdirectories, one level only: wiki/<topic>/<article>.md. Contains two special files:

• wiki/index.md — Global index. One row per article, grouped by topic, with link + summary + Updated date.
• wiki/log.md — Append-only operation log.

SKILL.md (this file) — Schema layer. Defines structure and workflow rules.

Templates live in references/ relative to this file. Read them when you need the exact format for raw files, articles, archive pages, or the index.

Initialization

Triggers only on the first Ingest. Check whether raw/ and wiki/ exist. Create only what is missing; never overwrite existing files:

• raw/ directory (with .gitkeep)
• wiki/ directory (with .gitkeep)
• wiki/index.md — heading # Knowledge Base Index, empty body
• wiki/log.md — heading # Wiki Log, empty body

If Query or Lint cannot find the wiki structure, tell the user: "Run an ingest first to initialize the wiki." Do not auto-create.

The Grounding Invariant

Every load-bearing fact in wiki/ — numbers, dates, direct quotes — exists verbatim in the raw/ files linked by that article's Raw field. Compile establishes this invariant (locate before you write); lint verifies it (scripts/check_evidence.py greps the high-signal literals — suffixed or large numbers, decimals, ISO dates, longer quotes — in the linked raws; the compile-time locate-before-write rule covers the rest). Because raw/ is immutable, a verified article stays verified; the script re-checks the whole wiki in seconds, so there is no incremental state to maintain.

Ingest

Fetch a source into raw/, then compile it into wiki/ — unless the source adds nothing new. Always fetch; whether to compile depends on the triage below.

Fetch (raw/)

1. Get the source content using whatever web or file tools your environment provides. If nothing can reach the source, ask the user to paste it directly.

2. Pick a topic directory. Check existing raw/ subdirectories first; reuse one if the topic is close enough. Create a new subdirectory only for genuinely distinct topics.

3. Save as raw/<topic>/YYYY-MM-DD-descriptive-slug.md.

   • Slug from source title, kebab-case, max 60 characters.
   • Published date unknown → omit the date prefix from the file name (e.g., descriptive-slug.md). The metadata Published field still appears; set it to Unknown.
   • If a file with the same name already exists, append a numeric suffix (e.g., descriptive-slug-2.md).
   • Include metadata header: source URL, collected date, published date.
   • Preserve original text. Clean formatting noise. Do not rewrite opinions.

   See references/raw-template.md for the exact format.

Triage

After saving the raw file and before editing wiki/, search wiki/ with the source's key entities and synonyms, then state the disposition:

• New — creates one or more new articles.
• Update — merges into existing article(s).
• Disputed — contradicts existing content; may combine with New or Update (see Compile for conflict annotation).
• No material — adds no knowledge beyond what the wiki already holds. Keep the raw file, log it (see Post-Ingest), and stop. Do not force an article out of a thin source.

New, Update, and Disputed may be combined. No material is exclusive.

Compile (wiki/)

Determine where the new content belongs:

• Same core thesis as existing article → Merge into that article. Add the new source to Sources/Raw. Update affected sections.
• New concept → Create a new article in the most relevant topic directory. Name the file after the concept, not the raw file.
• Spans multiple topics → Place in the most relevant directory. Add See Also cross-references to related articles elsewhere.

These are not mutually exclusive. A single source may warrant merging into one article while also creating a separate article for a distinct concept it introduces. In all cases, check for factual conflicts: if the new source contradicts existing content, mark the contested claims with a Status: Disputed block (see references/article-template.md). When the conflicting content lives in separate articles, mark both and cross-link them.

Source fidelity. Every number, date, and direct quote must be located in the raw file (grep or read) before it is written; write the value exactly as found — if the source says 42K, write 42K, not 42,000. Derived values (sums, deltas, counts you computed) must show their components so each component is findable in raw. If you cannot locate a value, do not write its exact form; drop it or state it without precision.

See references/article-template.md for article format. Key points:

• Sources field: author, organization, or publication name + date, semicolon-separated.
• Raw field: markdown links to raw/ files, semicolon-separated.
• Relative paths from wiki/<topic>/ use ../../raw/<topic>/<file>.md (two levels up to project root).

Cascade Updates

After the primary article, check for ripple effects. Do not rely on the index alone: search the full wiki for the source's key entities, aliases, and the claims it touches, then update every non-archive article whose content is materially affected. Each updated file gets its Updated date refreshed.

When the new source supersedes or contradicts an existing claim, keep the old claim for the record but mark it with a Status block (see references/article-template.md): Outdated when something newer replaces it, Disputed when sources disagree. Never silently rewrite history.

Archive pages are never cascade-updated (they are point-in-time snapshots).

Post-Ingest

Update wiki/index.md: add or update entries for every touched article. When adding a new topic section, include a one-line description. The Updated date reflects when the article's knowledge content last changed, not the file system timestamp. See references/index-template.md for format.

Append to wiki/log.md:

## [YYYY-MM-DD] ingest | <primary article title>
- Disposition: <New; Update; Disputed>
- Raw: <raw file path>
- Updated: <cascade-updated article title>

Omit - Updated: lines when no cascade updates occur. For No material, log and stop. Use a project-root-relative raw path (for example, raw/topic/file.md):

## [YYYY-MM-DD] ingest | no material: <project-root-relative raw file path>
- Disposition: No material

The exact no-material heading is the machine-readable inventory key; the Disposition line remains required for a complete human-readable log entry.

Research (multi-source ingest)

Use only when the user explicitly asks to research a topic or gather sources into the wiki. Ordinary knowledge questions go to Query, which never writes files.

1. Split the topic into a few angles. For each, search with a wide net — official names, abbreviations, and synonyms, not just the literal keywords.
2. For any core claim you expect to conclude, deliberately search the opposing side: failures, criticism, failed replications.
3. Save selected sources to raw/ as usual. Searching may run in parallel; compilation must not — compile one source at a time, because index.md, log.md, and cascade updates are shared state.

Query

Search the wiki and answer questions. Examples of triggers:

• "What do I know about X?"
• "Summarize everything related to Y"
• "Compare A and B based on my wiki"