By kfchou
Build and maintain an LLM-powered personal wiki: bootstrap at a path, ingest sources like papers/URLs/files into summarized pages with entity indexes, query for cited answers that flag gaps and offer new pages, audit for broken links/contradictions/staleness with fix reports, and revise pages with diffs/sources while updating links/index.
npx claudepluginhub kfchou/wiki-skills --plugin wiki-skillsUse when fact-checking a single wiki page against its cited sources — verifies that every footnote actually supports its claim and surfaces uncited factual claims. Run after ingesting a high-stakes page or any time you want confidence in one page's accuracy.
Use when adding a new source to a wiki — a paper, article, URL, file, transcript, or any document. One ingest may touch 10-15 wiki pages.
Use when bootstrapping a new personal wiki for any knowledge domain — research, codebase documentation, reading notes, competitive analysis, or any long-term knowledge accumulation project.
Use when auditing a wiki for health issues — contradictions between pages, orphan pages, broken cross-references, stale claims, missing pages, or coverage gaps. Run after every 5-10 ingests.
Use when asking a question against a personal wiki built with wiki-init and wiki-ingest. Do not answer from general knowledge — always read the wiki pages first.
Use when revising existing wiki pages because knowledge has changed, a new piece of information updates or contradicts existing content, or the user wants to directly edit wiki content with LLM assistance.
A Claude Code plugin implementing Karpathy's LLM Wiki pattern — a persistent, compounding knowledge base maintained by your LLM.
Instead of RAG (re-deriving answers from raw documents every time), this system builds and maintains a wiki: a structured, interlinked collection of markdown files that gets richer with every source you add and every question you ask.
The wiki format is remarkably similar to how the Claude Code Harness manages memory interally (read about it here). Since Claude is trained on the Claude Code Harness, that means Claude is already familiar with this structure. Unsure about other coding harnesses though.
/plugin marketplace add kfchou/wiki-skills
/plugin install wiki-skills@kfchou/wiki-skills
| Skill | Description |
|---|---|
wiki-init | Bootstrap a new wiki for any domain |
wiki-ingest | Add a source (paper, URL, file, transcript) to the wiki |
wiki-query | Ask a question against the wiki; optionally save the answer back |
wiki-lint | Health audit: contradictions, orphans, broken links, coverage gaps |
wiki-update | Revise existing pages when knowledge changes |
wiki-audit | Per-page citation audit: verify every footnote against its source, flag uncited claims |
<wiki-root>/
├── SCHEMA.md # Conventions + wiki root path (how skills find the wiki)
├── raw/ # Immutable source documents (you manage)
├── wiki/
│ ├── index.md # Content catalog — every page, one-line summary
│ ├── log.md # Append-only operation log
│ ├── overview.md # Evolving synthesis of everything known
│ └── pages/ # All wiki pages, flat, slug-named
└── assets/ # Images, PDFs, attachments
wiki-init → bootstrap a new wiki
wiki-ingest → add sources one at a time (repeat)
wiki-query → ask questions; save good answers back as pages
wiki-audit → fact-check a single page against its sources
wiki-lint → periodic health check (every 5-10 ingests)
wiki-update → revise pages when knowledge changes
wiki-ingest surfaces key takeaways and asks what to emphasize before writing anything. After creating a source page, it runs a backlink audit — scanning existing pages to add bidirectional links.wiki-query always reads the wiki (never answers from memory). Always offers to file the answer back as a new page with [[citations]].wiki-lint writes a severity-tiered report (🔴 errors / 🟡 warnings / 🔵 info) to wiki/pages/lint-<date>.md, offers concrete fixes, and logs unconditionally.wiki-audit fact-checks one page against its sources. Phase A flags uncited factual claims; Phase B dispatches one subagent per source in parallel to verify each footnote (quote citations are string-matched, [synthesis] citations are judged against the cited range). Writes a verdict report to wiki/pages/audit-<page>-<date>.md and offers concrete fixes.wiki-update always shows diffs before writing, always cites the source of new information, sweeps all pages for the same stale claim, and logs unconditionally.Works for any domain where you're accumulating knowledge over time:
Andrej Karpathy's LLM Wiki pattern (April 2026)
"The wiki keeps getting richer with every source you add and every question you ask."
PRs, issues, and ideas are welcomed.
MIT
A persistent LLM-maintained wiki for your knowledge base. Ingest sources, query your knowledge, and lint for consistency.
Share bugs, ideas, or general feedback.
LLM-maintained knowledge base skill — structured wiki with Obsidian, milestone-based source clustering, proactive write-back, and autonomous lint
Build and maintain an LLM-curated personal knowledge base in your project — Andrej Karpathy's LLM Wiki pattern, designed to scale to thousands of pages without becoming a context bottleneck. Now with an optional compiled graph layer for typed, provenance-backed relationships.
Quality assurance and linting tool for second brain knowledge systems ensuring consistency and structure. Knowledge base maintenance utility.
Compile scattered markdown knowledge files into a topic-based wiki. Non-destructive, Obsidian-compatible, with staged adoption.
Claude + Obsidian knowledge companion. Sets up a persistent, compounding wiki vault. Covers memory management, session notetaking, knowledge organization, and agent context across projects. Based on Andrej Karpathy's LLM Wiki pattern. Optional DragonScale Memory extension adds hierarchical log folds, deterministic page addresses, embedding-based semantic tiling lint, and boundary-first autoresearch topic selection.