菜单
Use when the user wants to control NotebookLM programmatically, import .docx into an Obsidian vault as atomic Zettelkasten notes, build a compiled LLM Wiki layer, or full-text-search the whole vault. Claude Code and Codex both supported.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when writing prompts for ComfyUI image or video models: choosing tags vs literary format, fixing composition and anatomy artifacts, picking samplers per LoRA, and writing short movement prompts for video. Neutral examples only.
Use when the user wants to run ComfyUI image/video generation on a rented remote GPU instead of paid API subscriptions: rent a GPU box, bootstrap the stack, connect the agent over SSH, drive workflows via the ComfyUI API, and deliver results to Telegram.
Use when writing or tuning ElevenLabs speech so it sounds alive, paced, and human; covers text shaping for Eleven Multilingual v2 and Eleven v3, plus API-based voice-setting updates from user feedback.
Use when the user needs their own VPN/proxy on a VPS to bypass regional blocks or DPI so research APIs and browsers work through a clean exit. One-script Xray VLESS+REALITY installer and client manager.
Use BEFORE installing any skill, setting, hook, or memory/instruction file into a user's setup. Detects the target environment, resolves real paths, asks when unsure, installs idempotently without overwriting, and verifies. Use when: "install this skill", "set this up for me", "add this to my agent", "configure my agent".
| id | obsidian-dataweave |
| name | ObsidianDataWeave: NotebookLM, Atomization, and LLM Wiki |
| summary | KISA's skill: turns Claude Code / Codex into a NotebookLM control panel + .docx import with Zettelkasten atomization + a compiled LLM Wiki layer + FTS5 memory across the whole Obsidian vault. All programmatic, all into your vault, with no external search dependencies. |
| type | skill |
| author | kisa |
| recommended | false |
| added | "2026-07-04T00:00:00.000Z" |
| tags | ["obsidian","notebooklm","zettelkasten","llm-wiki","fts5","memory","notes"] |
| source | https://github.com/howdeploy/ObsidianDataWeave |
| description | Use when the user wants to control NotebookLM programmatically, import .docx into an Obsidian vault as atomic Zettelkasten notes, build a compiled LLM Wiki layer, or full-text-search the whole vault. Claude Code and Codex both supported. |
| license | MIT |
Turns Claude Code and Codex into a full control panel for NotebookLM and your Obsidian
vault. You run deep research, manage sources, and pull notes out of notebooks — all with a
single natural-language command. In parallel, it imports .docx from Google Drive and
atomizes them into Zettelkasten notes with a MOC, tags, and wikilinks. On top — an isolated
LLM Wiki layer (a compiled base in the Karpathy style), and FTS5 memory — a full-text
index of the entire vault without a single external dependency. Public: howdeploy/ObsidianDataWeave, MIT.
Both clients are supported: Claude Code (skill) and Codex (AGENTS.md).
notebooklm-py as
a library (not the CLI — strictly one-shot, no retry duplication). Deep/fast research, source
management, dedup, note extraction, and atomization into the vault.[[wikilinks]].git clone https://github.com/howdeploy/ObsidianDataWeave.git
cd ObsidianDataWeave
bash install.sh --vault-path "/path/to/your/vault"
The installer checks for Python 3.10+, installs the dependencies (python-docx, pyyaml), creates
a config.toml with the vault path, registers the skill globally in ~/.claude/skills/obsidian-dataweave/,
and appends a block to ~/.claude/CLAUDE.md. After installation the skill works from any directory.
Modes: --mode claude (default), --mode codex (checks AGENTS.md), --mode local
(dependencies + config only). Updating is idempotent: git pull && bash install.sh.
After installation, just tell the agent in natural language:
| What to say | What happens |
|---|---|
process MyDocument.docx | Download → parse → atomize → write into the vault |
process note "Title" | Enrich or atomize an existing note |
process contacts "Contacts" | A note with contacts → personal cards + Networking MOC |
run research in notebook "<id>" "<query>" | Deep research in NotebookLM via the API |
clean duplicates in notebook "<id>" | Dedup sources in the notebook |
create wiki "<slug>" | Skeleton of a new LLM Wiki space (project/corpus, RU/EN) |
build wiki "<slug>" | Compile the raw material into pages (guard on [[wikilinks]]) |
search notes "<query>" | FTS5 search across the whole vault (bm25 + snippets) |
Processing modes: Enrich (short note → tags/links/expansion, 1→1), Atomize (long → atomic notes + MOC, 1→N), Contacts (contacts → cards + Networking MOC).
The third knowledge layer is a compiled wiki in the Karpathy style. It lives in the isolated folder
<vault>/LLM Wiki/<slug>/; atomic notes don't go there. It accumulates through an explicit
merge: existing [[wikilinks]] must be preserved, otherwise compile fails with
WIKI_LINKS_LOST. Two modes: project (fixed core pages: overview,
architecture, components, workflows, glossary…) and corpus (only entities/concepts grow).
python3 scripts/wiki_init.py demo --mode project --title "Demo Project"
python3 scripts/wiki_ingest.py demo path/to/article.md --kind articles
python3 scripts/wiki_compile.py demo --since-last-compile
python3 scripts/wiki_lint.py demo --strict
All scripts write through a single writer, vault_writer.py — the atomic pipelines and the wiki
share one write point.
The Google session is saved in ~/.notebooklm/storage_state.json — login is one-time.
python3 -m venv .venv
.venv/bin/python scripts/notebooklm_setup.py --skip-login
# in a SEPARATE terminal window (needs a real TTY):
.venv/bin/notebooklm login # log in to Google in Chromium → come back → ENTER
Then: process_notebook.py <notebook_id> (id is the last segment of the notebook URL),
research_notebook.py run "<id>" "<query>" --mode deep.
.docx), Claude Code or Codex.sqlite3) — no separate DBMS needed; doctor.py checks this.Related: NotebookLM++ — a Chrome extension for bulk-importing sources into NotebookLM (web, YouTube, playlists, PDF snapshots).