dossier

English · 中文

A portable Skills-format skill — works with Claude Code, Codex, ChatGPT, and other Skills-compatible agents. It lets your AI agent delegate citation-heavy research to Google NotebookLM, while keeping the corpus out of the agent's context window.

Why

When you ask Claude to "research topic X", three things normally go wrong:

1. Tokens burn fast. Pasting 20 PDFs into the conversation blows through context and the bill.
2. Answers hallucinate. Plausible-sounding but unsourced claims get mixed in with real facts.
3. Work doesn't compound. Next week, same topic, same sources — you start from scratch.

dossier splits the job:

Role	Who	What they do
📚 Corpus + retrieval	NotebookLM	Holds your sources forever. Answers questions with [1][2] citations that map back to source text, never extrapolating.
🧠 Orchestration	AI agent (Claude Code / Codex / ChatGPT / …)	Curates sources, decides when to delegate to the dossier, writes code, runs experiments, synthesizes.
🎯 Decisions	You	Confirm the shortlist of sources. Judge the final conclusion.

The result is a reusable, citation-backed corpus per research topic, where repeat questions cost cents instead of dollars and every factual claim traces back to source text.

When to use

• Literature reviews (multiple papers, cross-paper comparisons)
• Due diligence on companies, funds, sectors
• Regulatory / legal / compliance review across document sets
• Competitive intelligence
• Long-lived personal knowledge base (Obsidian / Readwise / meeting archives)
• Any task where citations are required and the same corpus will be queried repeatedly

When NOT to use

• One-off web lookups → use Claude directly
• Codebase exploration → your agent's native file tools (Read / Grep / etc.) are stronger
• Real-time data → NotebookLM is static
• Tiny corpora (< 5k tokens) → just paste into the prompt
• When latency matters more than citation quality → NotebookLM is ~3× slower than direct chat

How it works — the three routes

Every new dossier starts with one question: how are the sources going to get in?

Route	Best for	Who picks sources
A · NotebookLM auto-research	Quick overview, exploration	NotebookLM's own research agent (via web search)
B · Claude-curated ⭐	Authoritative sources needed, long-term reuse	Claude proposes a shortlist, you confirm, Claude imports
C · User-specified	You already know what you trust	You give the list, Claude just imports

The skill's working rule: never pick silently — present the tradeoff, let the user choose.

Install

1. Get the underlying tool

dossier builds on top of teng-lin/notebooklm-py (Python NotebookLM CLI). The skill itself will prompt you to run these on first use, but you can do it now:

pip install "notebooklm-py[browser]"
python3 -m playwright install chromium
python3 -m notebooklm login      # one-time Google OAuth via browser
python3 -m notebooklm skill install  # installs the low-level `notebooklm` skill

PATH note (macOS): pip install often puts the notebooklm binary in ~/Library/Python/<ver>/bin/ which isn't on $PATH. The skill handles this automatically by falling back to python3 -m notebooklm — you don't have to fix PATH.

2. Install this skill

dossier is in the Skills format — a single markdown file with YAML frontmatter that any Skills-compatible agent can load. Install path depends on your agent:

Agent	Install
Claude Code	git clone https://github.com/wangbooth/dossier ~/.claude/skills/dossier
Codex CLI	git clone https://github.com/wangbooth/dossier ~/.codex/skills/dossier
Multi-agent (Open Skills)	npx skills add wangbooth/dossier
ChatGPT / others	Check your agent's docs for its skill directory, then git clone the repo into it

If you already have the repo cloned somewhere, symlink instead of re-cloning:

ln -sfn /path/to/dossier <your-agent-skill-dir>/dossier

Codex skill-installer note: Codex's built-in installer currently has trouble with "root-as-skill" repositories like this one (SKILL.md sits at the repo root, not inside a subdirectory), because its git sparse-checkout path doesn't handle --path . gracefully. The plain git clone command above bypasses the installer and works reliably — just point it at ~/.codex/skills/dossier and restart Codex.

3. Jina Reader key — no setup needed upfront

Some sites (PubMed most notably) aggressively block NotebookLM's backend scraper. When that happens the skill falls back to Jina Reader, an LLM-friendly fetch proxy.

You don't need to do anything now. The first time a fetch actually needs Jina, Claude will ask you to paste a key (free from https://jina.ai, no credit card) and save it to ~/.config/dossier/config.json with chmod 600. From then on it's automatic — no export, no shell config, no key typed in a terminal.

Advanced users who want to preconfigure can either set $JINA_API_KEY (env var always wins) or write the config file directly:

// ~/.config/dossier/config.json
{ "jina_api_key": "jina_xxxxxxxx" }

Quick start

In your agent session (Claude Code, Codex, etc.), just describe your research:

"I want to research whether creatine is safe for the kidneys long-term."

Claude will detect the intent, load the skill, and walk you through the three-route choice. Typical flow for Route B (recommended):

1. Claude scans your existing notebooks — if a relevant dossier exists, it reuses it.
2. Otherwise Claude proposes a shortlist of authoritative sources (you say go / edit).
3. Claude imports sources, verifies they landed correctly (catches anti-bot failures), uses Jina Reader where needed.
4. You ask questions. Answers come back with [1][2] citations mapped to specific text spans in specific sources.
5. Next week, same topic, same dossier — skip steps 1–3.

Example

See examples/creatine-research.md for a real 8-source dossier built in ~15 minutes, with three live Q&A demonstrating the depth of cited answers.

Under the hood

The skill is pure prompt engineering — no code. It's a markdown file in the Skills format that teaches any compatible agent:

• When to trigger (description field)
• How to run the preflight (install check, auth check, PATH fallback)
• Which of the three routes to propose and how to branch
• The nine working rules (shortlist confirmation, citation resolution via --json, "outside source" layering, conversation clearing between subtopics, etc.)
• How to structure the output (## Dossier says / ## What I did / ## Conclusion / ## Not covered)

All commands route through the notebooklm CLI (from notebooklm-py).

Why notebooklm-py and not another client

We compared three active NotebookLM clients before picking one. Full decision record in research/client-comparison.md. TL;DR:

• teng-lin/notebooklm-py (11k★, Python) — chosen. Only client with a real create notebook command, best agent-integration story, PyPI-stable releases.
• icebear0828/notebooklm-client (181★, TypeScript) — missing create, npm package lags the main branch.
• PleasePrompto/notebooklm-mcp (2k★, TypeScript MCP) — chat-only, can't manage sources or dossiers; 4 months stale.

Known limitations

• Unofficial Google API. NotebookLM has no public API. notebooklm-py uses reverse-engineered endpoints that may break when Google changes their backend.
• Rate limits are real. PubMed specifically aggressively anti-bots cloud scrapers. The skill uses Jina Reader as a fallback.
• Session token decays. Re-run notebooklm login if auth errors start appearing (typically weeks-to-months).
• Source cap. Free NotebookLM tier = 50 sources per notebook, Pro = 300. Plan dossier scope.
• Generation commands are slow. notebooklm generate audio/video take 5–10 min each.

Contributing

Bug reports and PRs welcome. The skill is one markdown file (SKILL.md) — improvements typically come from running real dossiers and finding cases where the rules miss.

License

MIT — see LICENSE.