abz-karpathy-llm-wiki
CLAUDE.md1 starsChanged 6 months agoMITView on GitHub
Schema for the LLM Wiki. Read on every session. This is not a skill. It is the agent's default behavior, on every action. Skip it only when the user explicitly says "answer from general knowledge" or "don't check the wiki." Default to synthesis. Fewer pages with more cross-links beats many sparse pages.
In this repository3 other files it gives its agents.
- ingest-sourceSkill
- lint-wikiSkill
- save-sessionSkill
The file
Show its rules and text
We split this CLAUDE.md into the rules it gives an agent
13 rules. About 0 are common, 10 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈10
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 29
Before answering any question, drafting any document, or making any decision:
- Very specific: line 30
Read wiki/index.md to find what's available
- Distinctive: line 31
Read relevant synthesis pages and deep-dives
- Very specific: line 32
If the wiki has gaps for the question, fall back to raw/
- Distinctive: line 33
If you discover something durable while answering, file it back into the wiki
- Distinctive: line 41
Default to synthesis. Fewer pages with more cross-links beats many sparse pages.
- Distinctive: line 47
Every page uses YAML frontmatter:
- Distinctive: line 58
When ingestion finds a contradiction, flag it inline using this exact format:
- Distinctive: line 63
Do not silently overwrite.
- Very specific: line 70
Add your own SKILL.md files in .claude/skills/<skill-name>/ to extend.
- Distinctive: line 73
Read the wiki before doing work
- Distinctive: line 74
Update the wiki after doing work
- Distinctive: line 83
Create deep-dive pages by default — bias toward synthesis updates
≈ “Common” is an estimate: at least 1 in 5 of the 27,641 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
CLAUDE.md
Schema for the LLM Wiki. Read on every session.
What this is
A wiki built and maintained collaboratively by a human and an LLM agent. The wiki shapes itself around whatever sources you feed it — legal cases, medical research, code documentation, customer interviews, anything. Drop sources in raw/. The agent distills them into wiki/. You query the wiki before doing work.
Architecture — three layers
raw/— immutable sources. Files only move fromunprocessed/toprocessed/during ingestion; content is never modified.wiki/— agent-owned synthesis. Create, update, cross-link freely.CLAUDE.md— schema. Rules and pointers.
Directory layout
.
├── CLAUDE.md
├── .claude/skills/ skills (ingest-source, lint-wiki, save-session)
├── raw/
│ ├── unprocessed/ queue of new sources awaiting ingestion
│ ├── processed/ sources after ingestion
│ └── session-summary/ chronological session records
└── wiki/
├── index.md catalog of all wiki pages (one-line entries by category)
├── log.md append-only operations log (ingests, refactors, lint passes)
└── _templates/ page templates (copy when creating new pages)The query rule (always on)
Before answering any question, drafting any document, or making any decision:
- Read
wiki/index.mdto find what's available - Read relevant synthesis pages and deep-dives
- If the wiki has gaps for the question, fall back to
raw/ - If you discover something durable while answering, file it back into the wiki
This is not a skill. It is the agent's default behavior, on every action. Skip it only when the user explicitly says "answer from general knowledge" or "don't check the wiki."
Two output shapes — synthesis vs deep-dive
- Synthesis page — catalog-style, covers many entities of one type. Lives at
wiki/<category>.md. Default choice. - Deep-dive page — detailed coverage of one entity. Lives at
wiki/<category>/<name>.md. Created only when content has design-level depth that won't fit a synthesis row.
Default to synthesis. Fewer pages with more cross-links beats many sparse pages.
Page types
The agent invents page types from your sources. A legal wiki grows case, ruling, and precedent pages; a medical wiki grows study and cohort pages; an engineering wiki grows pipeline, table, and concept pages. When a structure repeats, the agent creates a template in wiki/_templates/. Default templates (concept, decision, plan, reference) are starting examples — extend, replace, or remove them.
Frontmatter conventions
Every page uses YAML frontmatter:
---
type: <page type>
status: draft | partial | complete | current
last_updated: YYYY-MM-DD
sources: [list of source paths]
---Conflict callout format (mandatory)
When ingestion finds a contradiction, flag it inline using this exact format:
⚠️ Conflict noted YYYY-MM-DD — <one-line title>
- Source A (<path>): <claim>
- Source B (<path>): <claim>
- Resolution: (empty until resolved) Do not silently overwrite. Do not pause to ask the user.
Skills
Procedures live in .claude/skills/:
ingest-source— process raw sources into the wiki (interactive or bulk)lint-wiki— audit wiki for stale pages, contradictions, orphanssave-session— capture session summary toraw/session-summary/Add your ownSKILL.mdfiles in.claude/skills/<skill-name>/to extend.
Workflow rules
- Read the wiki before doing work
- Update the wiki after doing work
- Distinguish raw from synthesis (raw stays in
raw/, distilled goes inwiki/) - Flag, don't overwrite (use conflict callouts)
- One thing per page (one entity, one concept, one decision)
- Standard markdown links (not
[[wikilinks]])
What this agent does NOT do
- Modify content of files in
raw/(read-only; only moved betweenunprocessed/andprocessed/) - Auto-fix structural wiki issues without confirmation
- Create deep-dive pages by default — bias toward synthesis updates
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

