llm-wiki

name	llm-wiki
description	Builds and maintains a persistent, interlinked Obsidian-compatible markdown knowledge wiki in a git repo via three operations — ingest a source into linked pages, answer a question from the accumulated wiki with citations, and lint the wiki for orphans, broken links, and contradictions. Use when the user asks to ingest or add a source into their wiki or knowledge base, init or grow such a wiki, ask what their wiki says about a topic, or lint and health-check it. Do not use for ad-hoc note-taking, code comments, or querying third-party or external wikis.
compatibility	Base wiki ops need git (+ Python 3 for optional scripts). Web scraping (fetch/crawl) is optional and works only in local Claude Code with network access — install Scrapling + trafilatura + a headless browser. Paper/PDF extraction needs pypdf. Scraping is unavailable on the Claude API sandbox (no network/installs) and unreliable on claude.ai.
allowed-tools	Bash(python3 *)
metadata	{"version":"0.3.0","author":"Charles Hoskinson"}

LLM Wiki

Knowledge is compiled once on ingest and kept current, so synthesis, cross-references, and contradictions already exist before any question is asked — the opposite of query-time RAG, where the model re-derives everything from raw chunks on every query.

The three layers

• Raw sources (raw/) — immutable files; capture operations may create them, then Claude reads them and never mutates them; the only ground truth.
• The wiki (wiki/) — LLM-generated, LLM-owned markdown pages derived entirely from raw sources and git history.
• The schema (CLAUDE.md) — tells Claude how this wiki is structured and which workflows to follow; co-evolves with the user.

Finding the wiki

A directory is a wiki if and only if it contains a CLAUDE.md whose first non-empty line carries the marker <!-- llm-wiki: v1 --> and a wiki/ folder. Resolution order: (1) explicit path provided by the user; (2) walk up from the current working directory to the first directory matching both conditions. The init operation creates a new wiki at any target path — nothing in this skill is hardcoded to a specific repo.

Operations (router)

Read the corresponding reference file before executing any operation:

• init — user says "set up a wiki here", "create a wiki", or similar → read references/operations/init.md
• ingest — user says "add this source", "ingest this", "process this file/paper/PDF into my wiki" (for a local file in raw/) → read references/operations/ingest.md
• scrape / crawl — user says "scrape this page", "crawl this site/section", "ingest this URL", "render JS", "fetch this link" → read references/operations/scrape.md (plain file ingest stays on the ingest entry above)
• query — user says "what does my wiki say about X", "look up X in my wiki" → read references/operations/query.md
• lint — user says "lint the wiki", "health-check the wiki", "find orphans/broken links" → read references/operations/lint.md

When the intent is ambiguous, ask the user which operation they want before proceeding.

Conventions

Before authoring any wiki page, read references/conventions.md for page types, link discipline, frontmatter schema, provenance markers, contradiction handling, naming/dedup rules, and MOC/index discipline. For Obsidian settings, Dataview queries, Web Clipper mapping, .gitignore setup, and optional Marp usage, read references/obsidian-setup.md.

Skill files

This skill expects these files to exist (open them as needed; if one is missing, the skill is incomplete):

• references/operations/{init,ingest,query,lint}.md — one playbook per operation
• references/operations/scrape.md — fetch/crawl playbook (optional, Claude Code only)
• references/conventions.md — page types, links, frontmatter, provenance, contradictions, naming/dedup, MOC/index
• references/obsidian-setup.md — Obsidian/Dataview/Web Clipper/.gitignore/Marp
• assets/page-templates/{source,concept,entity,synthesis,map}.md and assets/{schema,index,category-index,log}-template.md — templates used by init and ingest
• scripts/lint.py — optional mechanical lint accelerator (lint also works manually)
• scripts/scrape.py — optional web fetch/crawl script (Scrapling + trafilatura; Claude Code only)
• scripts/paper_vector.py — optional PDF paper fixture/extraction helper (pypdf)
• scripts/audit_checks.py — optional local mechanical audit runner

Changelog — v0.2.0 (additive): optional Scrapling+trafilatura web scrape/crawl on-ramp; base v1 operations unchanged. Changelog — v0.3.0 (additive): paper/PDF extraction helper, enforced scrape duplicate skips, response content-type quality gating, crawl link normalization, and repeatable audit checks.

Core rules (always)

• Never mutate captured files in raw/ — raw sources are immutable after capture; if a source needs annotation, create a source page in wiki/sources/ instead.
• Every claim on a wiki page carries an inline provenance marker with a locator or short quote anchor (e.g., Founded in 2019.^[from [[Source Page]] — "founded in 2019"]); a bare source link without a locator is a lint defect (UNSOURCED).
• Contradictions append, never overwrite — when a new source disagrees with an existing sourced claim, record both positions in a conflict callout and flag status: contested; resolution is a human decision.
• One atomic git commit per ingest — the source page, all entity/concept page edits, index updates, and the log entry go into a single commit referencing the raw/ file; this is the unit of undo and the backup mechanism.
• Every new page gets at least one outbound [[wikilink]] and is registered in its category _index.md and at least one MOC before the commit — orphan pages are a lint defect.