Reading with Ears — reading list & email triage

Primary Gmail search — union of all Gmail labels on enabled feeds (skip feeds with an empty gmail_labels list for the OR query only — they still get notebooks if mail is classified into them another way, e.g. manual triage):

• Build label:foo OR label:bar OR … for every distinct label across enabled feeds.
• Example when news / think / pro / healthcare / ai-everybody are all enabled:

gmail_search_messages(q="(label:newsletter/news OR label:newsletter/think OR label:newsletter/pro OR label:newsletter/healthcare OR label:newsletter/ai-everybody) after:YYYY/MM/DD")

Classification hint: If an email carries a label that appears in feed gmail_labels, default its read category to that feed’s notebook_category. The content classifier may still override when clearly wrong. Starred mail without any of those labels is a potential To-Do (report only).

Sender registry: Keep reading-with-ears/data/newsletter_sender_registry.json updated; add new senders under the right category when you adopt new labels.

Step 1: Determine Date Range & Fetch Emails

If the user says "today" or gives no date, use today only. If the user gives a range ("back to 3/15", "last few days", "this week"), use the earliest day as the start.

Primary search — use the OR-of-labels query built from enabled feeds’ gmail_labels (see above). If no enabled feed defines any label, report that misconfiguration and stop (or fall back to starred search only if the user explicitly asks).

Fallback — unlabeled starred emails (to-do triage):

gmail_search_messages(q="is:starred after:YYYY/MM/DD")

Run both searches when applicable. Deduplicate by messageId. Starred email that does not carry any newsletter label from the feeds config should be treated as a potential To-Do.

If no emails are found at all, tell the user and stop.

Step 2: Fetch Full Content

For each email returned, fetch the full message:

gmail_read_message(messageId=<id>)

Fetch ALL emails before doing anything else. This keeps tool calls front-loaded and avoids hitting the per-turn tool call limit mid-workflow.

HTML-only emails: Some emails return an empty plain-text body with a note to view the HTML version. When this happens:

• Use the subject line and snippet to create a minimal source entry
• Note in the title that the body was not accessible: "<subject> — <sender> [body unavailable]"
• Include whatever context is available from the snippet

Step 3: Classify Each Email

Classify every email as one of:

• One read bucket per enabled feed — use that feed’s notebook_category name in the triage table (e.g. 📰 News & Current Affairs, 🧠 Things to Think About, 💼 Professional Reading, 🎙️ AI is for Everybody, …).
• 📋 To-Do — anything requiring action, reply, decision, or follow-up

To-Do signals:

• Email does NOT have any label listed under enabled feeds’ gmail_labels
• Someone is waiting for a reply or decision
• Contains an action request, question, or deadline
• Sent emails (SENT label) that you starred to track — treat as follow-up items
• Subject patterns: "Re:", "Action required", "Please", "Can you", "Following up", receipts/invoices

Ambiguity rule:

Default to to-do if any action is implied, even loosely. Unlabeled starred emails default to to-do unless content is unambiguously a newsletter read.

Show triage table before acting:

Build one NotebookLM section per enabled feed that has at least one email (use notebook_category as the heading). Example shape when four feeds are enabled:

📧 Triage — [date range]

📰 News & Current Affairs (→ NotebookLM):
  • …

🧠 Things to Think About (→ NotebookLM):
  • …

💼 Professional Reading (→ NotebookLM):
  • …

🎙️ AI is for Everybody (→ NotebookLM):
  • …

📋 To-Do (→ report only — not NotebookLM, no task app):
  • "Re: Resume: Dave Holmes-Kinsella" — follow up with Olivier @ Ramp
  • "It's that time of year" — reply to birthday message from G
  • "Your receipt from Gamma #2239-6502" — file the receipt

Proceed? (or tell me what to reclassify)

Wait for confirmation. If the user reclassifies items, re-show the updated table and confirm again before acting.

Skip empty categories silently — if there are no professional reads, don't create that notebook. Do not call any task-manager MCP for to-dos.

Step 4: Route To-Reads → NotebookLM

Create one notebook per non-empty enabled feed (non-empty = at least one email in that feed’s read bucket). Assign nn by notebook_order among enabled feeds only: the lowest notebook_order gets 01, the next 02, etc. (pad to two digits).

Title format:

reading-list-YYYY-MM-DD-nn <notebook_category from feeds.json for that feed>

Example with four enabled feeds:

• reading-list-2026-04-05-01 📰 News & Current Affairs
• reading-list-2026-04-05-02 🧠 Things to Think About
• reading-list-2026-04-05-03 💼 Professional Reading
• reading-list-2026-04-05-04 🎙️ AI is for Everybody

If running for a date range, use the end date (today) in the name. If a notebook with today’s date already exists, increment the numeric suffix.

For each notebook:

Create it (substitute the feed’s notebook_category):

notebook_create(title="reading-list-YYYY-MM-DD-nn <notebook_category>")

Add each email as a text source:

source_add(
  notebook_id=<id>,
  source_type="text",
  title="<subject> — <sender> (<date>)",
  text="From: <sender>\nDate: <date>\n\n<body>",
  wait=True
)

• Truncate body to ~8,000 characters if very long
• For HTML-only emails with no body: use subject + snippet as the text

Generate Audio Overview after all sources are loaded — use that feed’s audio_focus_prompt from feeds.json (verbatim). If missing, fall back to a single generic ~12-minute insight-first prompt.

studio_create(
  notebook_id=<id>,
  artifact_type="audio",
  audio_format="deep_dive",
  audio_length="long",
  focus_prompt="<audio_focus_prompt from the matching feed>",
  confirm=True
)

Tool call budget awareness:

This workflow is tool-call-intensive. To stay within per-turn limits:

• Fetch all emails in Step 2 before creating any notebooks
• Create all notebooks before adding any sources
• Add sources notebook by notebook (all sources for notebook 1, then notebook 2, etc.)
• Start all audio generations last
• Run Step 5 (episode titling) after all audio is confirmed generated
• Run Step 6 (download) after all episodes are titled
• If the budget runs out mid-workflow, report exactly what was completed and what remains, so the user can continue in the next turn

Step 5: Title & Describe Each Audio Episode

After firing all audio generations, call studio_status on each notebook to confirm completion, then call notebook_describe to get the AI-generated summary. Use this to rename each artifact with a rich title, bullet-point key ideas, and sources line.

For each notebook:

studio_status(notebook_id=<id>)       # get artifact_id and confirm completed
notebook_describe(notebook_id=<id>)   # get AI summary to distill into bullets

Then rename:

studio_status(
  notebook_id=<id>,
  action="rename",
  artifact_id=<artifact_id>,
  new_title="<NotebookLM-generated title>\n\n• <key idea 1>\n• <key idea 2>\n• <key idea 3>\n\nSources: <Newsletter (topic)> · <Newsletter (topic)> · ..."
)

Titling rules:

• Keep NotebookLM's auto-generated title — it's usually excellent; don't replace it
• Bullets should be punchy and insight-first, not just descriptive ("X covers Y")
• Sources line format: Newsletter Name (topic shorthand) · Newsletter Name (topic) · ...
• 3 bullets minimum, 5 maximum — don't pad
• If audio is still in_progress when Step 5 runs, poll studio_status until completed

Example finished episode title:

How Power Profits From Manufactured Chaos

• AI companies are uniquely selling a product framed around existential risk and job displacement — yet people are buying it anyway
• The Iran war is functioning as a massive regressive global tax, hitting import-dependent nations hardest while the US gets off relatively easy
• Trump's second-term economic instability is self-generated, not inherited — a structural shift from his first term
• The White House governs through a repeating pattern: manufacture crisis → demand concessions → claim victory

Sources: Noahpinion (AI's worst sales pitch; Iran war economics) · Jonah Goldberg / The Dispatch (Two Trump economies) · The Atlantic Daily (Trump's bailout pattern)

Step 6: Download Audio Files

After all episodes are titled, download each completed audio file to the iCloud Personal Podcast folder so Phase 2 (Element.fm upload) can pick them up.

download_artifact(
  notebook_id=<id>,
  artifact_id=<artifact_id>,
  output_path="~/Library/Mobile Documents/com~apple~CloudDocs/Personal Podcast/YYYY-MM-DD-<slug>.mp3"
)

Slug mapping (from feeds.json)

Use each feed’s slug field for the filename: YYYY-MM-DD-<slug>.mp3 (e.g. news, think, professional, vital-signs, ai-everybody). Phase 2 (publish_episodes.py) uses the same slugs for enabled feeds.

• Skip download if the file already exists at that path (idempotency)
• If download fails, report the error and continue — Phase 2 will retry via manifest

Step 7: Report Back

✅ Done — [date range]

📚 NotebookLM — N notebooks created (one line per enabled feed that had mail):
  • "reading-list-YYYY-MM-DD-01 …" → K sources
    🎧 "<episode title>" → 📥 YYYY-MM-DD-<slug>.mp3
  • (repeat for each feed)

📋 To-dos (handle in Gmail — not exported to a task app):
  • "Re: Resume …" — follow up
  • …

Nothing was deleted or archived in Gmail.
Audio files downloaded to iCloud Personal Podcast — ready for Phase 2 (Element.fm upload).

Edge Cases

• No labeled emails in range: Fall back to starred search; tell user labels aren't set up if both return nothing
• All to-reads, no to-dos: Say so in the report; no task list section needed
• All to-dos, no to-reads: Skip NotebookLM; list to-dos in the report only
• HTML-only email body: Use subject + snippet, note body was unavailable
• Unlabeled starred email looks like a newsletter: Flag it in the triage table with a note, let user decide
• Gmail MCP unavailable: Tell user to check Settings → Connectors
• Duplicate notebook name: Increment nn suffix (01 → 02 → 03)
• User reclassifies: Re-show full triage table with changes, re-confirm before acting
• Tool call limit hit mid-run: Report what's done, what's pending, continue next turn
• Audio still in_progress at Step 5: Poll studio_status until completed, then rename
• Audio file already exists at download path: Skip download, note it in report-back
• Download fails: Report error, continue with remaining notebooks — Phase 2 will retry
• New newsletter not yet labeled: Add Gmail filter + label, add the label to the right feed’s gmail_labels in feeds.json, and note the gap to the user
• Enabled feed with empty gmail_labels: Primary search won’t catch mail; user must route content manually or add labels before relying on automation