Skillget

obsidian-llm-wiki

skill
Security Audit
Fail
Health Warn
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 6 GitHub stars
Code Fail
  • rm -rf — Recursive force deletion command in .claude/settings.local.json
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Compile notes into a linked Obsidian wiki at ingest time with Claude Code, following Karpathy's LLM Wiki idea instead of RAG-at-query.

README.md

Obsidian LLM Wiki

Compile notes at ingest time. Query a wiki that already exists.

This vault follows Andrej Karpathy's LLM Wiki. Traditional RAG retrieves from scratch on every question. Here, Claude Code compiles sources into a linked, stateful Obsidian wiki when you ingest them. Later queries read that compiled graph.

The human curates sources and decides. The model does bookkeeping: extract, link, index, flag conflicts.

本仓库按 Karpathy 的 LLM Wiki 思路，在摄入时用 Claude Code 把 raw/ 编译进 wiki/，而不是在查询时做 RAG。知识库正文用简体中文编写；本 README 以英文为准。

License: MIT

Layout

raw/                         inbox (content is not edited)
├── 01-articles/
├── 02-papers/
├── 03-transcripts/
├── 04-meeting_notes/
├── 09-archive/              processed sources (gitignored)
└── .processed.json          ingest ledger

wiki/                        compiled layer (the model owns this)
├── concepts/                frameworks, methods (TitleCase)
├── entities/                people, tools, products (TitleCase)
├── sources/                 source abstracts (kebab-case)
├── syntheses/               longer reports (kebab-case)
├── _conflicts/              conflicts that need a human
├── _lint/                   lint reports
├── index.md                 single index, grouped by type + tag cloud
└── log.md                   append-only operations log

assets/                      images, PDFs
CLAUDE.md                    vault rules for the agent
.claude/skills/              agent skills

raw/ is immutable text. After a successful ingest, the source file moves to raw/09-archive/. wiki/ is the compiled, stateful output.

Setup

  1. Install Obsidian.
  2. Install Claude Code.
  3. Clone and open the folder as a vault:
git clone https://github.com/levi-qiao/obsidian-llm-wiki.git
cd obsidian-llm-wiki

In Obsidian: Open folder as vault, then choose this directory.

claude

Obsidian CLI is optional. It needs a Catalyst license and a running Obsidian app. /ingest, /query, and /lint work without it.

Commands

/ingest                         scan raw/ for unprocessed files
/ingest raw/01-articles/x.md    ingest one file
/query what do my notes say about X?
/lint                           health check

/sync commits wiki/ when you want a Git snapshot. It is optional.

Ingest

Reads new files in raw/, skips anything already in .processed.json, and compiles them into wiki/.

The model extracts a thesis, entities, concepts, and a few tags. It updates or creates pages, writes a source abstract, refreshes index.md and log.md, then archives the source. It does not edit the source text.

Query

Reads wiki/index.md, opens a small set of pages, and answers with [[wikilink]] citations. High-value answers can be saved as a synthesis if you say so.

Three modes:

  • /query <question> — precise
  • /query --explore <topic> — map coverage
  • /query --relate <A> <B> — shortest path between two pages

Lint

Checks index drift, missing frontmatter, dead links, orphan pages, one-off tags, and marked conflicts. Safe index and frontmatter fixes run automatically. The rest is a report.

Skills

Fourteen skills ship in .claude/skills/.

Wiki loop:

  • ingest — compile raw/ into wiki/, then archive
  • query — index-first answers with wikilinks
  • lint — health check

Obsidian (Kepano):

  • defuddle — clean a web page to Markdown
  • obsidian-cli — vault operations from the CLI
  • obsidian-bases — Bases views
  • obsidian-markdown — Obsidian-flavored Markdown
  • json-canvas — JSON Canvas files

Diagrams:

  • obsidian-canvas-creator — Canvas layouts
  • mermaid-visualizer — Mermaid from text
  • excalidraw-diagram — Excalidraw drawings

Study:

  • tutor-setup — turn a document or repo into a study vault
  • tutor — quizzes and progress
  • scholar-skill — layered paper reading

Naming and frontmatter

  • Entities and concepts: TitleCase (ClaudeCode.md, 注意力机制.md)
  • Sources and syntheses: kebab-case (摘要-attention-is-all-you-need.md)
  • Every page needs a ## 关联连接 section with [[wikilinks]]. No orphans.
---
type: concept | entity | source | synthesis
tags: [tag1, tag2, tag3]    # 3–5; reuse tags from the cloud
source_id: stable-id        # optional
created: 2026-04-29
updated: 2026-04-29
---

Tags emerge from content. Prefer the existing cloud over new labels.

What Git tracks

Tracked: wiki/, pending files under raw/01–04, assets/, .claude/skills/, CLAUDE.md, and most .obsidian config.

Ignored: Obsidian workspace files, plugin data.json, .claude/memory/, and raw/09-archive/*.

Caveats

  • Token cost. Each command ships a lot of context. Watch the API bill.
  • Hallucination. Compiled pages can be wrong. Run /lint and read the pages that matter.
  • Scale. This fits a few hundred good sources. A much larger corpus needs real search infrastructure.

License

MIT. See LICENSE.

Links

Reviews (0)

No results found