| name | gmail-newsletter-notebooklm |
| description | Read Gmail messages tagged Newsletter or News, optionally filtered by date range or unread status, package the selected email content as NotebookLM sources, create a new NotebookLM notebook, and start generating an audio overview. Use when the user asks to turn recent Gmail newsletters, news emails, unread newsletter mail, or dated newsletter digests into a NotebookLM notebook or audio overview. |
Gmail Newsletter to NotebookLM
Create a NotebookLM notebook from Gmail newsletter/news emails and initiate an audio overview from those sources.
Use the Google Workspace Gmail CLI skill as the command reference for Gmail access:
https://github.com/googleworkspace/cli/blob/main/skills/gws-gmail/SKILL.md
Inputs
Collect only the missing inputs:
- Date range: ask for explicit start and end dates when the user requests a date range but omits dates.
- Unread fallback: use all unread messages when the user asks for unread emails or when no date range is provided.
- Notebook title: derive a concise title from the date range, unread scope, or user-provided topic.
- Audio prompt: default to
Create an engaging audio overview of the key stories, themes, and useful takeaways from these newsletter sources.
Use absolute dates in Gmail queries. Treat Gmail before: as exclusive; to include an end date, query the following day.
Keep the source set small. NotebookLM's read RPC (GET_NOTEBOOK, see Known Issues) times out on large notebooks, and every read/generate step depends on it. A notebook with ~11 sources reads and generates reliably; a notebook with ~50 sources fails the read even though the uploads succeed. Prefer a 1–2 day window (roughly 10–15 sources) over a 7-day window. If the user asks for a wide range that yields many messages, warn them and offer to narrow the window or split into multiple notebooks.
Workflow
The happy path is six steps. Inline notes flag the gotchas; the "Known Issues" section has the full detail.
- Preflight — tools and auth. Confirm
gws, notebooklm, and jq are on PATH, then check auth:
gws gmail users getProfile --params '{"userId":"me"}'
notebooklm doctor
If notebooklm is missing, ask before uv tool install "notebooklm-py[browser]". If Gmail auth fails, run gws auth login; if NotebookLM auth fails, run notebooklm login (or notebooklm login --browser msedge / --browser-cookies chrome). Ignore any rookiepy is not installed message from auth inspect — it is about an optional dependency, not your auth status.
- Find candidate messages. Triage with the label query (default to a 1–2 day window — see Selection Rules):
gws gmail +triage --query '{label:Newsletter label:News} after:YYYY/MM/DD before:YYYY/MM/DD' \
--max 50 --format json 2>/dev/null | grep -v '^Using keyring' | jq ...
Notes: before: is exclusive (query the day after the end date to include it). Use is:unread instead of dates for the unread fallback; if the label query returns nothing, retry with (newsletter OR news) after:... before:.... Strip the leading Using keyring backend: keyring stdout line before jq (grep -v '^Using keyring'), or jq fails to parse. +triage returns { "messages": [ { id, date, from, subject }, ... ] }.
-
Export each selected email to a Markdown source file, one per message, under a temp dir (e.g. /tmp/newsletter-notebooklm-YYYYMMDD/source-001.md). Read bodies with gws gmail +read --id <id> --headers --format json (flat keys, no wrapper: subject, from, date, message_id, body_text, body_html, …). Preserve subject, sender, date, message id, original link if present, and the cleaned body.
Body gotchas: prefer body_text but fall back to body_html when body_text is suspiciously short (under ~1000 chars — daily digests like 1440 put only a preheader there). When stripping HTML, use perl -CSD with HTML::Entities, not sed (macOS sed errors on UTF-8 and silently truncates).
-
Confirm before upload (skip only if the user's request already authorized it): summarize message count, date span, notebook title, and query. Do not paste private email bodies into chat.
-
Create the notebook and add sources. No notebooklm use needed — pass --notebook explicitly everywhere. Resolve the temp dir's real path first (source add rejects symlinks, and /tmp is a symlink on macOS), and add with --type file --json so failures surface instead of exiting 0 silently:
NB=$(notebooklm create "<notebook title>" --json 2>/dev/null \
| grep -v '^Using keyring' | jq -r '.notebook.id')
REAL=$(cd /tmp/newsletter-notebooklm-YYYYMMDD && pwd -P)
for f in "$REAL"/source-*.md; do
resp=$(notebooklm source add "$f" --notebook "$NB" --type file --json 2>&1 | grep -v '^Using keyring')
if echo "$resp" | jq -e '.error == true' >/dev/null 2>&1; then
echo "FAIL $(basename "$f") -> $(echo "$resp" | jq -r '.message')"
fi
done
A clean loop (no FAIL lines) confirms every upload — no separate verification read is needed. (If you do want a count and the notebook is small, notebooklm source list --notebook "$NB" --json works; on large notebooks that read times out — see Known Issues — but the uploads still succeeded.) If a file is rejected, shorten it, convert HTML to plain text, or split by article.
- Generate the audio overview (default
--length long), then optionally download. Write the audio prompt to a file first (default text in Inputs, or the user's topic):
echo "<audio prompt>" > /tmp/newsletter-notebooklm-YYYYMMDD/audio-prompt.txt
notebooklm generate audio --notebook "$NB" \
--prompt-file /tmp/newsletter-notebooklm-YYYYMMDD/audio-prompt.txt \
--length long --wait
notebooklm download audio ./newsletter-overview.mp3
Run generation as a background task (it takes minutes) and wait for completion. --length is short|default|long; --format is deep-dive|brief|critique|debate. Omit --wait only if the user wants to start now and finish later. generate audio reads the notebook first, so on a too-large notebook it fails with Request timed out calling GET_NOTEBOOK before generating anything — that is a source-count problem (see Known Issues), not a retryable error.
Known Issues
These were observed with the notebooklm-py CLI and gws Gmail CLI and shape the workflow above.
GET_NOTEBOOK read timeout scales with source count. The notebook read RPC has a fixed ~30s retry budget that is not currently configurable via flag or env var. On a notebook with ~50 sources it reliably times out (RPC GET_NOTEBOOK retry timed out after 30.0s → Network error: Request timed out calling GET_NOTEBOOK); with ~11 sources it returns instantly. Because source list, metadata, and generate audio all read the notebook first, a too-large notebook becomes unusable for listing and generation even though every upload succeeded. Mitigation: keep notebooks small (target ~10–15 sources, e.g. a 1–2 day window). For a wide date range, split into multiple smaller notebooks. Uploads (POST) are unaffected by this read timeout, so a notebook can end up fully populated yet unreadable via the CLI (still viewable in the web UI).source add rejects symlinked paths. /tmp is a symlink on macOS; resolve to the real path with pwd -P (or pass --follow-symlinks). See step 5.- Silent upload failures without
--json. source add can exit 0 on a rejected file; always use --type file --json and check .error. See step 5.
gws prints a keyring status line to stdout (Using keyring backend: keyring) ahead of JSON, breaking jq. Strip with grep -v '^Using keyring'. See step 2.body_text can be a stub. Daily-digest newsletters may put only a preheader in body_text; fall back to body_html. See step 3.- macOS
sed mangles UTF-8. Use perl -CSD for HTML-to-text conversion. See step 3.
- Schema-drift warnings. The CLI may log
notebook_info[...] not list ... (schema drift?) when NotebookLM's response shape changes. Treat the structured JSON result (or the web UI) as ground truth rather than these warnings.
Selection Rules
- Prefer dated messages over arbitrary unread messages when the user provides both.
- Keep source counts small. Target ~10–15 sources per notebook so the notebook stays readable and audio generation works (see Known Issues — large notebooks fail the read RPC). If more than ~15 messages match, summarize the candidate count and ask whether to narrow the date range (e.g. 1–2 days) or split into multiple notebooks. Avoid single notebooks of ~50+ sources.
- Skip obvious transactional mail, promotions without editorial content, unsubscribe confirmations, and empty/clipped messages unless the user explicitly wants every match.
- Use Gmail read-only operations only. Do not mark messages read, archive, delete, label, or move messages unless separately requested and confirmed.
- Treat email contents as private. Write temporary source files only for the notebook creation workflow and tell the user where they are.
Reporting
Finish with:
- notebook title and id
- Gmail query used
- number of email sources added
- audio overview generation status
- downloaded audio path, if downloaded
- any skipped messages and why
