dossier
Skill1 starsChanged 7 months agoMITView on GitHub
Use when investigating a topic that benefits from a reusable, citation-backed NotebookLM corpus—literature reviews, due diligence, regulatory/legal research, competitive intelligence, long-term personal knowledge queries, or any repeated deep-dive across the same body of material. Not for one-off web search, single-document reads, or code exploration.
- Reads credentials
- Installs packages
12 other projects use a version of this.
Adopted by
Most recently taken up by
- DeepExtrema/Tesseracttheir own version, first seen Nov 16, 2026
- yuanhaorannnnnn/agent-skillstheir own version, first seen Nov 1, 2026
- koffih/claude-skill-dossiertheir own version, first seen Nov 1, 2026
- eins78/agent-skillstheir own version, first seen Oct 30, 2026
- JohnnyMcGee/claude-skillstheir own version, first seen Oct 17, 2026
- samuel-neveugall/client-dossiertheir own version, first seen Oct 17, 2026
Other versions12 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 12
- pnp/sharepoint-skills★ 142same name
- Hassaan146/claude-skills★ 18same name
- DeepExtrema/Tesseract★ 4same name
- eins78/agent-skills★ 2same name
- Hussain-Elfalah/Agentic-Skills★ 2same name
- mahiunfiltered/claude-code-skills★ 1same name
- samuel-neveugall/client-dossier★ 1same name
- dacrypt/tesla-claude-plugin★ 0same name
- JohnnyMcGee/claude-skills★ 0same name
- koffih/claude-skill-dossier★ 0same name
2 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
The file
Show its rules and text
We split this skill into the rules it gives an agent
15 rules. 0 are common, 4 are distinctive, and 11 are very specific to this project.
0
Common
Most of the 11 edited versions we read keep these.
4
Distinctive
This author's own choices. Worth a look.
11
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 25
When speed matters over depth — NotebookLM is ~3× slower than direct chat
- Very specific: line 60
If the notebooklm binary lands outside $PATH (typical on macOS where pip install writes to ~/Library/Python/<ver>/bin/), don't try to patch …
- Very specific: line 72
First branch — does a dossier on this topic already exist?: Run $NBLM list before anything else.
- Distinctive: line 77
Always present the choice to the user before spending time.
- Very specific: line 109
All commands use the active dossier (set by $NBLM use <id>).
- Very specific: line 125
--titleonly works for inline text: When adding a file with source add /path/file.md --title "...", the title is silently dropped and the … - Very specific: line 153
Import with the same hygiene rules as Route B (check titles after add, use Jina/DOI fallback for anti-bot sites, rename files since --title …
- Very specific: line 163
Don't ask for a Jina key during preflight — most dossiers never need one.
- Very specific: line 205
After saving, do not echo the key back.
- Very specific: line 226
If fulltext is tiny or the guide's keywords are off-topic, the source is effectively bad even if Ask answers look plausible (Ask is just …
- Distinctive: line 239
Add an inline summary note when no scrapable form exists — books you own offline, paywalled papers, Cloudflare-walled blog posts:
- Very specific: line 263
Do a refresh pass before a new round of Q&A on a dossier that's been idle for weeks — cheaper than realizing mid-research that answers are …
- Very specific: line 267
Check for existing dossier first: Before creating a new one, run $NBLM list and scan for a dossier on the same or adjacent topic.
- Very specific: line 273
One dossier per topic: Don't mix "creatine research" and "Q4 earnings analysis" in one notebook — cross-contamination ruins citation …
- Distinctive: line 281
When delivering results to the user, structure as:
“Common” is measured: kept by at least 6 in 10 of the 11 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Dossier
Open a topic-scoped NotebookLM corpus (the "dossier"), let Claude orchestrate curation and reasoning, let NotebookLM serve citation-backed answers from within the dossier. The dossier persists—future questions on the same topic reuse it.
When to use
- Reading 5+ papers/reports on one topic
- Pre-IPO / pre-deal due diligence on a company or sector
- Legal, regulatory, or compliance review across document sets
- Competitive analysis across multiple products/companies
- Long-term personal knowledge base queries (Obsidian/Readwise exports, meeting archives)
- Any task where answers must be traceable to source text
When NOT to use
- One-off questions answerable from general knowledge or single web page — just ask Claude directly
- Codebase exploration or "jump to definition" — your agent's native file tools (Read/Grep/etc.) are stronger
- Real-time / freshly changing information — NotebookLM is static
- Tiny corpora (< 5k tokens) — fitting in Claude's context is simpler
- When speed matters over depth — NotebookLM is ~3× slower than direct chat
Setup (first run in a machine)
Preflight has 4 steps. Only step 1 is read-only and always safe to auto-run. Steps 2–4 have side effects (disk writes, browser popups, skill directory changes) — ask the user before each one.
Step 1 — Detect (auto, read-only)
# What CLI form is reachable?
if command -v notebooklm >/dev/null 2>&1; then
NBLM="notebooklm"
elif python3 -m notebooklm --help >/dev/null 2>&1; then
NBLM="python3 -m notebooklm" # pip-installed but scripts dir not on PATH — common on macOS
else
NBLM="" # not installed
fi
# Does the existing session work?
if [ -n "$NBLM" ] && $NBLM auth check 2>&1 | grep -q "Authentication is valid."; then AUTH_OK=1; else AUTH_OK=0; fi
# Note: matches the final verdict line — only appears when every sub-check
# passed. Matching "✓ pass" alone would false-positive when e.g. only
# "Storage exists" passed but SID/token checks failed.Step 2 — Install (only if $NBLM is empty) — ASK FIRST
Tell the user: "I don't see notebooklm-py on this machine. May I install it with pip install 'notebooklm-py[browser]' plus Playwright Chromium (~150MB)?". Wait for an explicit yes, then:
pip install "notebooklm-py[browser]"
python3 -m playwright install chromium
NBLM="python3 -m notebooklm"If the notebooklm binary lands outside $PATH (typical on macOS where pip install writes to ~/Library/Python/<ver>/bin/), don't try to patch PATH yourself — just report it and continue using the python3 -m notebooklm form everywhere. It works identically.
Step 3 — Authenticate (only if AUTH_OK=0) — ASK FIRST, OPENS BROWSER
$NBLM login opens Chrome for Google OAuth. Tell the user it'll pop a browser window before running.
Step 4 — Install the low-level notebooklm skill (once per machine) — ASK FIRST
$NBLM skill install writes to ~/.claude/skills/notebooklm/ — it gives Claude direct knowledge of the raw CLI commands, which complements this dossier skill. Low risk but still a filesystem change; tell the user, then run.
Routes — pick the right entry point
First branch — does a dossier on this topic already exist? Run $NBLM list before anything else.
- Yes (same or adjacent topic) → Route D (query / top-up / hygiene). Most time on any long-running topic is spent here, not in fresh builds.
- No → pick one of A/B/C below based on source-quality requirements. Default: Route B.
Always present the choice to the user before spending time.
| Route | When | Setup time | Claude tokens | Durability |
| A. NotebookLM deep research | Quick overview, one-shot curiosity | ~5 min | near-zero | ephemeral (dossier stays but seeded by auto-research) |
| B. Claude-curated sources ⭐ | Noisy topic, long-term reuse, authoritative-source requirement | 15–30 min | moderate (fetching/filtering) | persistent |
| C. User-specified sources | Strict source whitelist (specific authors, databases) | depends on list | moderate | persistent |
| D. Existing dossier | Continuing research on a topic that already has a dossier | near-zero | low (mostly Q&A, maybe small top-up) | persistent |
Route A — NotebookLM auto-research
$NBLM create "dossier-<topic-slug>"
$NBLM use <notebook-id>
$NBLM source add-research "<search query>" --mode deepNotebookLM's own research agent picks sources, imports them, and indexes. Claude then asks questions. Source quality is opaque — good for exploration, risky for decisions.
Route B — Claude-curated (recommended default)
Steps:
- Claude searches authoritative sources (e.g. PubMed, Cochrane, ISSN, SEC EDGAR, primary sources for the topic — not random blogs or SEO pages)
- Claude drafts a source shortlist with a one-line justification per source
- User confirms or edits the list — never skip this step
- Claude imports confirmed sources one by one (source type auto-detected from content):
All commands use the active dossier (set by$NBLM source add "https://..." # URL $NBLM source add ./paper.pdf # local file $NBLM source add "pasted text" --title "My note" # inline text$NBLM use <id>). Add-n <id>to override. - Verify import quality (see caveats below), fix any bad imports
- Claude starts asking questions with
$NBLM ask --json "..."
Import caveats — read before step 4:
- Check every import. Run
$NBLM source listafter a batch. If a title readsChecking your browser - reCAPTCHA,Access denied, or similar, NotebookLM's backend hit an anti-bot wall. One retry is fine; if still blocked, abandon the direct URL and go to the Jina fallback below — don't loop on retries. - Never batch-parallelize PubMed URLs.
pubmed.ncbi.nlm.nih.govtriggers reCAPTCHA extremely reliably under any concurrency, and even a 3-second serial delay isn't enough. Go straight to Jina Reader + DOI for PubMed. - Jina Reader fallback — when a URL hits anti-bot walls, fetch it via Jina's proxy and ingest as a file. See "Jina key handling" below for how to obtain
$JINA_API_KEY:curl -sL -H "Authorization: Bearer $JINA_API_KEY" \ "https://r.jina.ai/<original-url>" -o /tmp/page.md $NBLM source add /tmp/page.md $NBLM source rename <source-id> "Author Year — descriptive title" - Prefer DOI over PubMed URL.
https://doi.org/<doi>redirects to the publisher. If it's MDPI / BMC / Frontiers / PLOS (open access), you'll get full text, which indexes 5–10× more useful content than a PubMed abstract page. Get the DOI from NCBI esummary:curl "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esummary.fcgi?db=pubmed&id=<PMID>&retmode=json". --titleonly works for inline text. When adding a file withsource add /path/file.md --title "...", the title is silently dropped and the filename becomes the title. Always follow a file import with$NBLM source rename <id> "..."to fix.
Route C — User-specified
User provides the source list (files, URLs, or "only these authors"). Same commands and same caveats as Route B, just skip steps 1–3.
Route D — Existing dossier (query / top-up / hygiene)
Pick this whenever $NBLM list shows a dossier on the same or adjacent topic. This is the default path for continuing research. It's not a build mode — A/B/C are for filling an empty dossier; D is for everything you do after.
Open the dossier, then branch into one of three sub-paths:
$NBLM list # find the dossier
$NBLM use <id> # activate it
$NBLM source list # remind yourself what's in hereD1. Query only
The dossier already covers the question — you just need an answer. Go straight to Ask, but respect Working rule 10 (split or -s-scope wide questions before they time out).
D2. Top-up (incremental source add)
Q&A revealed a specific gap — one angle, a few missing sources. Do not re-run Route B end-to-end. Do a mini-curation instead:
- Name the gap in one sentence (e.g. "no primary source on
<X>'s 2024 pricing change") - Propose 1–5 new sources with one-line justifications, user confirms
- Import with the same hygiene rules as Route B (check titles after add, use Jina/DOI fallback for anti-bot sites, rename files since
--titleis silently dropped on file imports) $NBLM source wait <id>if the next ask is time-sensitive — otherwise NotebookLM may answer before the new source is indexed- Re-ask the gap question. If still unanswered, treat it as a real data problem (no public source exists, dossier source cap hit) rather than a curation problem — don't keep adding sources of diminishing quality
D3. Hygiene pass
Symptoms that trigger this: citations point to garbage titles (reCAPTCHA / access-denied), Ask returns "outside source material" for things you know are covered, duplicates after multiple top-up rounds, or the dossier has been idle for weeks. See Source hygiene (ongoing) below.
Jina key handling (on-demand, only when a Jina fallback is actually needed)
Don't ask for a Jina key during preflight — most dossiers never need one. Only when a specific import hits an anti-bot wall and needs r.jina.ai does this flow run.
Step 1 — Resolve the key (auto, read-only)
Try sources in this order; take the first non-empty:
CONFIG_FILE="${XDG_CONFIG_HOME:-$HOME/.config}/dossier/config.json"
JINA_API_KEY="${JINA_API_KEY:-}" # 1. honor env var if already exported
if [ -z "$JINA_API_KEY" ] && [ -f "$CONFIG_FILE" ]; then
JINA_API_KEY=$(python3 -c "import json; print(json.load(open('$CONFIG_FILE')).get('jina_api_key',''))" 2>/dev/null)
fi
export JINA_API_KEYStep 2 — If still empty, ask the user (don't save silently)
Say something like:
"I need to fetch
<URL>through Jina Reader because NotebookLM's direct fetch hit a reCAPTCHA wall. I don't see a Jina API key saved on this machine. You can get a free one at https://jina.ai (no credit card; click the big API key box on the homepage). Paste it in your next reply and I'll save it to~/.config/dossier/config.jsonwithchmod 600so only you can read it. From then on it's automatic."
Step 3 — When the user pastes the key, save it
Pass the key through a heredoc, not argv (argv ends up in shell history):
mkdir -p "${XDG_CONFIG_HOME:-$HOME/.config}/dossier"
python3 <<'PY'
import json, os
p = os.path.expanduser(os.environ.get("XDG_CONFIG_HOME", "~/.config") + "/dossier/config.json")
p = os.path.expanduser(p)
existing = {}
if os.path.exists(p):
existing = json.load(open(p))
existing["jina_api_key"] = """<PASTE-USER-KEY-HERE>"""
os.makedirs(os.path.dirname(p), exist_ok=True)
json.dump(existing, open(p, "w"), indent=2)
os.chmod(p, 0o600)
print(f"Saved to {p}")
PYAfter saving, do not echo the key back. Confirm with just: "Saved to ~/.config/dossier/config.json. Continuing with the fetch."
CI/CD / one-off overrides
Setting JINA_API_KEY in the environment always wins over the config file. Useful for:
- CI jobs that inject the key via secrets
- Testing with a temporary key without overwriting the saved one
Source hygiene (ongoing)
Route B's "Import caveats" covers hygiene at import time. This section covers what to do after sources are in, because a dossier degrades over time — accumulated bad imports, URL rot, duplicates from top-up rounds.
Spot-check what was actually indexed
A clean title in source list doesn't guarantee the content was ingested — NotebookLM occasionally indexes just a cookie banner or a single page of a multi-page article. When an Ask answer seems suspiciously thin on a source you added, verify:
$NBLM source fulltext <id> | head -100 # first 100 lines of actually-indexed text
$NBLM source guide <id> # AI summary + auto keywords — reveals what NotebookLM "sees"If fulltext is tiny or the guide's keywords are off-topic, the source is effectively bad even if Ask answers look plausible (Ask is just leaning on other sources). Fix it before more Q&A.
Delete-and-replace bad sources
$NBLM source delete -y <id> # by id (supports prefix match)
$NBLM source delete-by-title "Exact Title" # errors if multiple matchAlways replace, don't just delete — otherwise future Ask silently loses that angle. Replacement options:
- Refetch via Jina Reader (see Route B caveats) if original hit anti-bot
- Ingest the DOI / publisher URL instead of PubMed
- Add an inline summary note when no scrapable form exists — books you own offline, paywalled papers, Cloudflare-walled blog posts:
Tag the title with$NBLM source add "<your summary text>" --title "<Author Year — topic> (notes)"(notes)or(summary)so later citations are unambiguous about what's primary vs. paraphrased.
De-duplicate after top-ups
Top-ups often re-add the same paper via different URLs (PubMed ↔ DOI ↔ arXiv). Run $NBLM source list periodically and look for:
- Same author / year / title across different rows
- An abstract-only version and a full-text version of the same paper — keep full-text, delete abstract
- Same URL hostname cluster on overlapping topics (often indicates blog-post drift into the corpus)
Refresh stale URL sources
URL and Drive sources can go stale when the upstream page changes:
if $NBLM source stale <id>; then
$NBLM source refresh <id>
fiDo a refresh pass before a new round of Q&A on a dossier that's been idle for weeks — cheaper than realizing mid-research that answers are from a cached old version.
Working rules
- Check for existing dossier first. Before creating a new one, run
$NBLM listand scan for a dossier on the same or adjacent topic. If found, go to Route D — do not rebuild. - Ask dossier before answering from memory. Any claim about content inside the corpus goes through
$NBLM ask "...". Do not answer from LLM parametric knowledge when the dossier should know. - Never paste source text into the Claude conversation. The whole point is to keep bulk material out of Claude's context. If you find yourself about to quote a long PDF, stop — ask the dossier instead.
- Preserve citations, and resolve them. The text answer contains
[1][2]markers — those are only meaningful paired with thereferencesarray fromask --json. Always run$NBLM ask --json ...and map each citation number to itssource_id+cited_textin the output. Bare[1]without resolution is misleading. - If the dossier can't answer, say "dossier has no coverage" — do not extrapolate. Listing unanswered questions is more useful than inventing answers.
- Curation is not optional in Route B/C. Show the user the shortlist before importing. Bad sources poison every future query.
- One dossier per topic. Don't mix "creatine research" and "Q4 earnings analysis" in one notebook — cross-contamination ruins citation quality.
- Clear conversation when changing subtopic.
askdefaults to continuing the last conversation (is_follow_up=true). When the user switches from one angle (e.g., safety) to another (e.g., dosage), run$NBLM clearfirst so the new question isn't biased by prior context. - Honor NotebookLM's "outside the source material" layer. When the answer contains phrases like "资料范围之外" / "outside the provided sources" / "not in the documents", NotebookLM is signaling speculation vs. sourced fact. Preserve this split in your output: put sourced content under
## Dossier says, put speculation under## Not covered by dossierwith an explicit "unverified" flag. Never merge the two. - Split wide Asks; don't retry them. NotebookLM can time out or return shallow answers when one
askspans too many sources (observed failures around 9+ sources in practice). When a question naturally crosses many sources, either split by angle (ask 2–3 narrower questions and synthesize) or use-s <source_id>(repeatable) to scope each ask to 1–5 sources. Never retry the same wide question more than once — split instead. - Keep hygiene up to date — don't run on a dirty corpus. When you spot a bad source mid-session (garbage title, suspiciously short
source fulltext, duplicate of another source), stop and fix it before continuing. One bad source poisons every Ask that touches it, and hygiene debt compounds over top-up rounds. See Source hygiene (ongoing).
Output format
When delivering results to the user, structure as:
## Dossier says
<facts from $NBLM ask, citations preserved>
## What I did
<commands run, files written, experiments executed>
## Conclusion
<your answer to the user's original question>
## Not covered by dossier
<questions the dossier couldn't answer — flagged for human follow-up>Quick command reference
$NBLM list # list existing dossiers
$NBLM create "dossier-<slug>" --json # new empty dossier, emit id as JSON
$NBLM use <id> # set active dossier for session
$NBLM status # show active dossier
$NBLM auth check # verify Google session still valid
$NBLM source add "https://..." # import URL (auto-detected)
$NBLM source add ./paper.pdf # import local file (auto-detected)
$NBLM source add-research "<query>" --mode deep # Route A auto-import
$NBLM source list # list sources in active dossier
$NBLM ask --json "<question>" # chat with active dossier — ALWAYS use --json to get citation→source mapping
$NBLM ask --json "..." -n <id> # chat with explicit dossier
$NBLM ask --json "..." -s <src1> -s <src2> # source-limited ask (repeatable -s) — use to dodge timeouts and force focus
$NBLM source wait <id> # block until a newly-added source finishes indexing
$NBLM source fulltext <id> # dump NotebookLM's indexed text — verify what was actually ingested
$NBLM source guide <id> # AI summary + keywords for a source — quick sanity check
$NBLM source stale <id> # exit 0 if URL/Drive source needs refresh, 1 if fresh (shell-scriptable)
$NBLM source refresh <id> # re-fetch a stale URL/Drive source
$NBLM source delete -y <id> # delete source (use -y to skip confirmation)
$NBLM source delete-by-title "..." # delete by exact title (errors if multiple match)
$NBLM generate audio --wait # podcast from dossier
$NBLM generate report --template study_guide # markdown report
$NBLM summary # AI-generated overview of dossierFull CLI reference: $NBLM --help or $NBLM <cmd> --help.
Known limitations
- Unofficial API.
notebooklm-pyuses reverse-engineered Google endpoints. Commands may break after Google backend changes — reinstall latest version if errors appear. - Session expires. Re-run
$NBLM loginif calls start returning auth errors. - Route A source quality is opaque. NotebookLM picks sources via web search; they may include blogs, wikis, or promotional pages. For high-stakes decisions, prefer Route B.
- Notebook source cap. Free tier 50 sources, Pro tier 300. Plan dossier scope accordingly.
- Generation commands are slow.
audio/videotake 5–10 min. Always use--waitwhen you need the artifact. - Single-ask source-count ceiling. In practice, one
askover ~9+ sources frequently times out or returns shallow synthesis. Use-sto scope or split the query (see Working rule 10).
Red flags — stop and reconsider
- About to paste a whole PDF into the Claude conversation → use
$NBLM source add --fileinstead - About to answer a factual question about the corpus from memory → use
$NBLM askinstead - Creating a new dossier because you forgot the old one existed →
$NBLM listfirst - Running Route B/C from scratch when
$NBLM listalready shows a relevant dossier → switch to Route D - Skipping the shortlist-confirmation step in Route B to "save time" → bad sources will cost 10× more time later
- Asking a question that spans 6+ sources without
-sscoping or splitting → timeout / shallow-answer risk - Citations point to a source whose
source fulltextis 50 lines of cookie banner → hygiene pass before more Q&A - Dossier returns an answer without citations → something's wrong, re-check the query or source indexing
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

