chubbyskills

ChubbySkills local ingest pack

Use this skill when the user wants LLMwiki to behave like a local NotebookLM / NetbookLM over their own feeds and saved content.

LLMwiki's role:

• Own the vault path, search, citations, graph, memory, review, and promotion workflow.
• Query saved Markdown through vault.search, query.unified, vault.read, memory.*, and AI-Output review tools.
• Keep governance boundaries: raw source, compiled wiki, agent output, reviewed decisions.

ChubbySkills' role:

• Own platform-specific capture and transcription.
• Convert links from Douyin, Bilibili, Xiaohongshu, WeChat, X/Twitter, podcasts, YouTube, Weibo, Zhihu, TikTok, and related feeds into Markdown assets.
• Use its own optional dependencies for video/audio transcription, subtitles, images, and enrichment.

Source project: https://github.com/chubbyguan/chubbyskills.

Integration model

Do not copy large upstream dependencies into the LLMwiki MCP server.

Instead:

1. Install LLMwiki normally.
2. Install upstream ChubbySkills separately when the user wants multi-channel capture.
3. Point both systems at the same local vault:

export VAULT_MIND_VAULT_PATH=/path/to/your/vault
export VAULT_DIR="$VAULT_MIND_VAULT_PATH"

4. Let ChubbySkills write captured Markdown into the vault.
5. Let LLMwiki search, cite, enrich, review, and promote the result.

Recommended vault layout

raw/                         # LLMwiki raw compiler sources
素材库/                       # ChubbySkills raw captured material
wiki/                        # compiled / structured knowledge
00-Inbox/AI-Output/<actor>/  # agent-authored review candidates
00-Inbox/Agent-Memory/<actor>/
20-Decisions/
30-Architecture/
40-Runbooks/

If the user's vault already has a different Chinese folder structure, preserve it. Do not force migration.

Skill routing

Need	Upstream ChubbySkills skill	LLMwiki follow-up
Douyin video	douyin-transcribe	query.unified, vault.writeAIOutput
Bilibili video	bilibili-transcribe	query.unified, concept compilation
YouTube video	youtube-transcribe	bilingual summary, tags, AI-Output
Podcast	podcast-transcribe	learning notes, memory handoff
WeChat article	wechat-article-ingest	source-backed wiki page
Xiaohongshu	xiaohongshu-ingest	image/video routing, hook analysis
X/Twitter single tweet	x-ingest	raw Markdown ingest
X/Twitter high-signal browser capture	x-to-obsidian	Web Clipper capture, LLMwiki search
Any captured note	content-enrich	summary, tags, value judgment
Whole vault management	knowledge-base-management	LLMwiki doctor, health, MCP query

Helper script

This skill bundles a tiny helper that prints safe install and environment commands:

python3 scripts/chubbyskills_bridge.py --help
python3 scripts/chubbyskills_bridge.py list
python3 scripts/chubbyskills_bridge.py plan --vault /path/to/vault --skills bilibili-transcribe podcast-transcribe

The helper does not install or clone anything by default. Treat it as a reproducible checklist.

Operating procedure

1. Identify the source URL or content channel.
2. Pick the upstream skill using the routing table.
3. If upstream ChubbySkills is not installed, run the helper plan command and show the user the commands.
4. Run upstream capture only when the local environment is ready.
5. Confirm the Markdown note exists under the vault.
6. Use LLMwiki MCP tools to search and summarize the saved note.
7. File agent-authored synthesis under 00-Inbox/AI-Output/<actor>/.
8. Promote durable conclusions only after review.

Good prompts

/chubbyskills install the Bilibili and podcast ingest path for my LLMwiki vault
/chubbyskills turn this Bilibili video into a searchable vault note, then summarize it
/chubbyskills capture these WeChat articles into 素材库 and build a wiki index
/chubbyskills make this vault work like a local NotebookLM over my saved feeds

Safety rules

• Respect source platform terms and rate limits.
• Do not bypass login, paywalls, private content, or deleted content.
• Do not hardcode cookies, API keys, or tokens into vault notes.
• Keep heavy transcription dependencies outside LLMwiki's MCP bundle.
• Prefer source-preserving Markdown over lossy summaries.
• Never claim "local NotebookLM" completeness unless retrieval was tested against the saved notes.