wikillm

LLM-maintained knowledge bases powered by Claude Code and Obsidian.

Drop sources in raw/, and Claude compiles them into a cross-referenced wiki — no vector stores, no embeddings, no RAG pipeline. Just markdown.

"I think there is room here for an incredible new product instead of a hacky collection of scripts."

This project is inspired by Andrej Karpathy's post on using LLMs to build personal knowledge bases — collecting raw sources, compiling them into a markdown wiki, using Obsidian as the visual layer, and having the LLM do all the writing, linking, linting, and Q&A. wikillm turns that workflow into a single command.

What you get

Run npx wikillm once, answer a few questions, and you'll have:

• A vault folder (.kb/ in a project, ~/.wikillm/<name>/ for personal) with raw/ for sources, wiki/ for compiled articles, and an Obsidian-ready workspace config
• A tailored CLAUDE.md inside the vault that tells Claude Code how to operate the knowledge base (directory layout, conventions, automation schedule)
• Seven Claude Code skills for ingesting, querying, linting, and presenting your KB
• Optional scheduled automation that ingests new sources and lints the wiki on a cron

From then on, you drop files into raw/, run /wikillm:ingest, and ask /wikillm:query whatever you want. Claude reads the compiled wiki — not the raw sources — so queries are fast and answers are consistent.

Requirements

• Claude Code (required) — the CLI tool that runs the skills
• Node.js 18+ (for npx wikillm)
• Obsidian (recommended) — for graph view, indexed search, and the Obsidian CLI. Enable via Settings → General → Command line interface once installed.
• Marp CLI (optional) — only if you want the slides feature for generating decks from wiki content

Setup

wikillm has two install steps: the Claude Code plugin (gives you /wikillm:* skills) and the npx wikillm CLI (scaffolds a vault). You need both.

1. Install the plugin in Claude Code

In an active Claude Code session, run these three slash commands:

/plugin marketplace add Berkay2002/wikillm
/plugin install wikillm@wikillm
/reload-plugins

They're idempotent — running them a second time is a safe no-op. After /reload-plugins, you should see wikillm:query, wikillm:ingest, wikillm:lint, and the other skills in your skills list. Invoke /wikillm:using-wikillm for the full orientation.

2. Scaffold a vault with npx wikillm

From a terminal, in whatever folder should contain the knowledge base:

# Personal KB (standalone vault)
cd ~ && npx wikillm

# Project KB (creates .kb/ in the current repo)
cd path/to/your/project && npx wikillm

The wizard asks for:

• Name — what to call the KB (e.g. research, nexus-kb)
• Kind — personal or project
• Mode — solo (full automation, you maintain it) or team (.kb/ is shared via git, manual ingestion)
• Features — slides, reports, visualizations, web-clipper (each has a description in the wizard)
• Domain — optional free-text (e.g. machine learning research); used to seed tag vocabulary in the generated CLAUDE.md
• Automation — daily/weekly ingest + weekly lint (solo modes only; requires Claude Desktop running when the schedule fires)

When it finishes you'll have a vault folder with this structure:

<vault>/
├── CLAUDE.md              ← schema for this KB (LLM-owned, tailored to your config)
├── raw/                   ← drop your sources here
│   └── assets/
├── wiki/                  ← compiled articles (created by /wikillm:ingest)
│   └── _index/
│       ├── INDEX.md       ← content catalog
│       ├── TAGS.md        ← tag directory
│       ├── SOURCES.md     ← raw/ → wiki/ provenance map
│       ├── RECENT.md      ← last 20 changes
│       └── LOG.md         ← operation log
├── outputs/               ← generated slides/reports/visualizations (if enabled)
└── .obsidian/             ← Obsidian workspace config

First-run walkthrough

After setup, the typical first run looks like this:

1. Drop reference material into raw/ — articles, PDFs, markdown docs, clipped webpages, whatever you want compiled. Keep meaningful subfolders (raw/<topic>/) so later lookups are easier.
2. Compile the wiki — in Claude Code, run /wikillm:ingest. It detects unprocessed files, extracts concepts, writes cross-linked articles, updates the indices, and commits one git commit per source file. Bulk imports (3+ files) dispatch parallel workers.
3. Query it — /wikillm:query "your question". Claude reads the compiled wiki — not the raw sources — and picks an appropriate output format (inline answer, report, slide deck, or visualization).
4. (Optional) Open in Obsidian for graph view, backlinks, and visual search: obsidian open vault="<your-kb-name>".

Modes

| Mode | Vault location | Automation | Git policy | |---|---|---|---| | Personal | ~/.wikillm/<name>/ | Daily ingest, weekly lint (scheduled) | Private vault, solo commits | | Project Solo | <repo>/.kb/ | Daily ingest, weekly lint (scheduled) | raw/ committed, .obsidian/ workspace gitignored | | Project Team | <repo>/.kb/ | Manual only — pull before ingesting | raw/ + wiki/ committed, coordinate pushes |

Scheduled automation requires Claude Desktop to be running when the schedule fires. For terminal-only workflows, stick with manual /wikillm:ingest runs.

Skills

All seven skills are installed via the Claude Code plugin:

| Skill | What it does | |-------|-------------| | /wikillm:using-wikillm | Orientation — how to use the plugin, detect a .kb/ vault, integrate with a project's existing CLAUDE.md | | /wikillm:query | Answer questions against the wiki. Picks an appropriate output format (inline, report, slides, visualization) | | /wikillm:ingest | Compile new files from raw/ into wiki articles with cross-references. Dispatches parallel workers for bulk imports | | /wikillm:lint | Health check — broken wikilinks, orphan pages, missing frontmatter, contradictions, stale claims | | /wikillm:obsidian-cli | Control Obsidian from the terminal — search, read, write, graph queries. Keeps Obsidian's index and graph in sync | | /wikillm:marp-cli | Generate slide decks (PDF/PPTX/HTML) from wiki content | | /wikillm:generate-schema | Regenerate a vault's CLAUDE.md from scratch. Use only if you customize config after init |

CLI commands

| Command | What it does | |---------|-------------| | npx wikillm / npx wikillm init | Interactive vault setup | | npx wikillm doctor | Check dependencies and vault health (run from inside a vault) | | npx wikillm update | Print the slash commands to update the plugin in Claude Code | | npx wikillm --help | Show available commands |

Integrating a KB with an existing project's CLAUDE.md

npx wikillm never touches your project's root CLAUDE.md — the vault's schema lives at <vault>/CLAUDE.md only, so your project's own instructions are never clobbered.

After setup, in Claude Code run /wikillm:using-wikillm and ask it to wire the KB into your project's existing CLAUDE.md. The skill contains the canonical integration pattern (dev-time KBs, team mode, KB-aware subagent rules) and will produce the right snippet for your project without you having to paste anything by hand.

How it works

1. You drop sources (articles, PDFs, notes, clipped webpages) into raw/
2. /wikillm:ingest reads them, extracts concepts, writes cross-linked wiki articles, updates four indices (INDEX, TAGS, SOURCES, RECENT), and commits each source as a separate git commit
3. For bulk imports (3+ sources), ingest dispatches parallel ingest-worker subagents with a concept-ownership table so workers don't duplicate articles, then a reconciliation pass merges and cross-links their output
4. Obsidian gives you graph view, indexed search, and backlinks over the compiled wiki
5. /wikillm:query answers questions by reading wiki articles directly — not by grepping raw sources
6. /wikillm:lint runs periodically (or on-demand) to catch broken wikilinks, orphan pages, missing frontmatter, and contradictions

This is not RAG. Knowledge is compiled once into the wiki and kept current via ingest and lint — not re-derived from raw documents on every query. The cross-references are already there. The contradictions have already been flagged. The synthesis already reflects everything that's been read. Querying is fast because the hard work happens at compile time.

Troubleshooting

"/wikillm:* commands don't appear after install" Run /reload-plugins to reload the plugin list. If they still don't appear, confirm the marketplace was added with /plugin marketplace list — you should see wikillm in the output.

"Obsidian CLI not detected" (or "binary found but not responding") Two independent things need to be true: (1) Obsidian desktop must be installed and running, (2) Settings → General → Command line interface must be enabled in the running instance. Until both hold, wikillm falls back to direct file tools, which works fine — you just miss graph view and indexed search.

Note that obsidian --version succeeds even when the desktop app is closed (it only tests that the binary is installed). The real liveness check is obsidian vaults — if it errors or hangs, Obsidian isn't running and you should start the desktop app before retrying.

"search:context (or other colon subcommands) exits 127 on Windows" On Git Bash for Windows, Obsidian CLI subcommands whose names contain a colon — search:context, property:set, daily:append, base:query, dev:*, and similar — can fail with exit code 127 due to how Git Bash parses argv[1]. Workarounds: run from PowerShell or CMD (colons survive argv parsing correctly there), or use the non-colon equivalent when one exists (plain obsidian search for most lookups). wikillm's skills already fall back to Grep/Read for operations that can't be worked around, so day-to-day flows aren't blocked — this mainly matters if you're calling the CLI directly from scripts.

"CLAUDE.md wasn't written to my vault after npx wikillm" This was a silent-failure bug in 0.1.0 where the init flow shelled out to a subprocess that couldn't write the file. Fixed in 0.2.0 — the generator now runs in-process and verifies the file after writing. If you're on 0.1.0, upgrade with npx wikillm@latest and re-run in the same directory (it'll prompt to overwrite the existing vault scaffold).

"Scheduled automation didn't run" Scheduled ingest and lint require Claude Desktop to be running when the schedule fires. If you work exclusively from the terminal, run /wikillm:ingest manually after adding sources and skip the automation opt-in during npx wikillm.

"I want to query an old version of an article" raw/ is immutable and the wiki is under git — use git log on wiki/<article>.md to see the history, or git show <rev>:wiki/<article>.md to read an older version.

License

MIT