Built to stop losing context between Claude sessions. The problem isn’t that Claude forgets — it’s that maintaining a structured knowledge base across hundreds of sessions by hand drifts and breaks down.

The plugin gives Claude three slash commands over an Obsidian vault:

• /llm-wiki:ingest captures any source (URL, YouTube, local file, pasted text) with citation discipline and idempotency guards
• /llm-wiki:query routes questions through a hot cache and domain indexes before reading individual pages
• /llm-wiki:lint runs twelve health checks with six auto-fixable categories

v0.2 in progress — more command options and a write-schema skill that bootstraps the CLAUDE.md wiki schema from scratch.

/plugin install llm-wiki@jordanchoi

Built as the transcript layer for llm-wiki-plugin. I needed reliable YouTube access regardless of whether a video had manual subtitles, what language it was in, or whether YouTube had recently changed its internals — so I forked an existing MCP server and replaced the hand-rolled scraper with yt-dlp. That moves the maintenance burden to a community of thousands instead of me.

Returns transcript lines with optional timestamps, language fallback, SponsorBlock-based ad-stripping (opt-in), and video metadata.

npx -y @jordanchoi/mcp-server-youtube-transcript
# requires yt-dlp on PATH: brew install yt-dlp