An Agent Skill That Compiles Notes Instead of Re-deriving Them

What it does

llm-wiki-skill is an agent skill—essentially a bundle of prompts, scripts, and conventions—that plugs into Claude Code, Codex, OpenClaw, or Hermes. It ingests PDFs, text, and web content, then compiles them into a structured local markdown wiki with bidirectional links, entity pages, and an offline HTML knowledge graph. The core pitch is “compile once, maintain continuously”: the agent organizes and caches knowledge so it persists across sessions rather than re-analyzing raw sources on every query.

The interesting bit

The offline graph viewer is styled like a Chinese landscape painting (“digital mountain-water”) and runs entirely in a self-contained HTML file—no server required. More practically, the system bakes in skepticism: every fact gets a confidence label like EXTRACTED or UNVERIFIED, and a SHA256 cache prevents weak models from duplicating work. It is essentially glue code, but it glues together the boring parts—ingest validation, hook injection, and cross-reference health checks—that keep a personal wiki from rotting.

Key highlights

• Offline “digital mountain-water” knowledge graph with search, filtering, and community clustering in a single HTML file.
• Bidirectional [[wiki links]], automatic entity/topic pages, and Obsidian-compatible local markdown.
• Confidence annotations (EXTRACTED, INFERRED, AMBIGUOUS, UNVERIFIED) flag which claims need verification.
• Session-start hooks automatically inject the knowledge base context into agent conversations.
• SHA256 deduplication and “self-healing” cache to avoid redundant ingestion, even with weaker models.

Caveats

• Optional web extractors (WeChat, YouTube, X/Twitter, Zhihu) require extra tooling like uv or bun, and some need logged-in browser sessions via Chrome DevTools.
• Windows PowerShell 5.1 users must deal with GB2312/ASCII encoding issues; the project provides a PowerShell wrapper or recommends upgrading to PS7.

Verdict

Worth a look if you already live inside an AI agent and want your research to accumulate instead of evaporate. Skip it if you are looking for a managed cloud solution or don’t want to babysit local scripts and dependencies.