Skip to main content
Status: Brainstorm Phase: Phase 5 (AI Media) | Tier: Pro / Studio

​
Overview

Google’s NotebookLM proved something profound: people will listen to two AI voices discuss their own data for twenty minutes and enjoy every second of it. The format works because conversation is the oldest and most natural way humans absorb information. A dialogue between two hosts who disagree, riff on each other, and get excited about findings is inherently more engaging than a monologue — even when both hosts are AI. AI Podcasts bring this format to cannabis intelligence. Two AI hosts — one analytical, one experiential — discuss your latest order, your terpene trends, your monthly consumption patterns, or a head-to-head comparison of two strains. They reference your data naturally: “So John just picked up Blue Dream for the third time. That is a pattern.” “Yeah, and look at the Myrcene content — this batch is way higher than the last one. He might notice a different body feel.” The conversation flows like two knowledgeable friends reviewing your purchase, not two robots reading a report. This is the AI media format with the strongest unit economics and the highest surprise-and-delight factor. Podcast generation is fast (under 60 seconds), costs are low (0.10−0.10-0.50 per episode), and the format itself is proven. NotebookLM has millions of users generating podcast episodes from PDFs and documents. We are generating them from structured cannabis intelligence data, which is both more personal and more interesting than a PDF. The ritual potential is enormous. “My monthly High IQ podcast dropped” becomes a sentence that subscribers actually say. Monthly recap episodes create recurring engagement touchpoints that feel earned (you generated this content through your purchases) rather than manufactured (here is a push notification because we need your DAU).

​
What It Does

​
Episode Formats

Two hosts discuss a specific strain from the user’s latest order. They cover the terpene profile, effects forecast, High Family classification, how it compares to similar strains, and any notable medical research. If the user has tried this strain before, the hosts reference their history: “This is your third time buying this. Your notes from last time said it was ‘perfect for Sunday mornings.’”
“Your Month in Cannabis” review episode. The hosts walk through all orders from the past month — what was purchased, how the terpene profile shifted compared to previous months, which High Families dominated, spending patterns, and standout strains. This is the flagship format because it creates a monthly ritual.
The hosts look at the user’s current active stash and discuss the composition: “You’ve got six strains right now, and four of them are Myrcene-dominant. That is a LOT of body sedation in one stash.” They suggest what is missing, what pairs well together, and what to try next.
A head-to-head comparison episode. The hosts debate two strains the user owns or is considering: effects, terpenes, use cases, High Family, potency. One host argues for each strain. Entertaining, informative, and useful for purchase decisions.
Any article from the TIWIH blog can become a podcast episode. The hosts discuss the article’s key points, add color commentary, and relate the topic back to cannabis science. This extends the content marketing reach of every blog post.

​
Host Personalities

The two hosts have distinct roles that create natural conversational dynamics:
HostRoleVoiceStyle
Host A — “The Analyst”Data-driven, precise, slightly nerdyWarm but measuredLeads with terpene data, research citations, classification details
Host B — “The Enthusiast”Experience-driven, curious, relatableEnergetic, casualAsks “but what does that actually feel like?”, relates to personal anecdotes, makes jokes
The contrast between them is what makes the conversation listenable. Host A says “the Caryophyllene content at 1.2% suggests anti-inflammatory properties.” Host B says “okay but real talk — is this the strain you reach for after leg day or what?” Neither is wrong. Together they cover both the science and the experience.

​
User Value

The “aha moment” is hearing two AI voices discuss YOUR specific strains by name, reference YOUR purchase history, and have a genuine-sounding conversation about YOUR cannabis data. The personalization transforms a novelty (“AI can make podcasts”) into a utility (“this is the most convenient way to understand my order”).

​
How It Works

1

Trigger Event

A podcast is triggered by one of: new research report generated (auto-generates Strain Deep Dive), month rollover (auto-generates Monthly Recap), user taps “Generate Podcast” on their stash or a strain comparison, or a new blog article is published.
2

Data Assembly

The system assembles all relevant structured data for the episode context: strain profiles, terpene arrays, High Family classifications, user history, order details, notes, and comparison metrics. This is NOT the prose report text — it is the raw structured data that enables natural conversation.
3

Dialogue Script Generation

An LLM (Claude) receives the structured data and a detailed prompt specifying the episode format, host personalities, conversation arc, and user context. The output is a dialogue script with speaker labels, timing cues, emphasis markers, and natural conversational elements (interruptions, laughter markers, “hmm” moments). The script is 800-2,000 words depending on episode length.
4

Text-to-Speech Rendering

Each host’s lines are sent to a TTS API (ElevenLabs) with distinct voice profiles. The voices are configured for conversational pacing — not audiobook narration. Emphasis markers from the script adjust pitch, speed, and volume. Word-level timestamps are returned for potential visual synchronization.
5

Audio Assembly

Individual audio clips are assembled into a continuous conversation with natural overlap timings, brief pauses between speakers, and ambient room tone to prevent clinical silence. A subtle intro jingle (2-3 seconds) and branded outro are prepended and appended. The final audio is exported as MP3 (128kbps, stereo).
6

Delivery

The episode is uploaded to CDN storage. A push notification announces it: “Your new podcast is ready.” The episode appears in the podcast section of the app with a play button, episode title, duration, and scrubbing progress bar. A 15-second preview is auto-extracted for the free tier upgrade hook.

​
Technical Approach

​
Pipeline Architecture

Report Data (JSON)
       │
       ▼
┌──────────────────┐
│  Dialogue Script  │  LLM generates two-speaker dialogue
│  Generation       │  from structured strain/order data
└──────────────────┘
       │
       ▼
┌──────────────────┐
│  Voice Rendering  │  TTS API renders each speaker
│  (Host A + B)     │  with distinct voice profiles
└──────────────────┘
       │
       ▼
┌──────────────────┐
│  Audio Assembly   │  Merge voices, add timing,
│  & Post-process   │  intro/outro, normalize levels
└──────────────────┘
       │
       ▼
┌──────────────────┐
│  CDN Upload +     │  Store MP3, send push notification,
│  Notification     │  extract 15-sec preview clip
└──────────────────┘

​
Dialogue Script Quality

The script generation prompt is the most critical piece. A bad prompt produces stilted dialogue that sounds like two people reading alternating paragraphs of a report. A good prompt produces conversation that sounds like two friends at a dispensary counter. Key prompt techniques:
  • Give hosts distinct knowledge gaps — Host B should not know everything, so Host A has things to explain
  • Include disagreement moments — “I actually think the Limonene is doing more work here than you’re giving it credit for”
  • Reference specific data points naturally — “1.2% Caryophyllene” not “the Caryophyllene percentage”
  • Include reaction beats — “Wait, really?” “Okay that’s actually interesting” “I didn’t expect that”
  • End with a takeaway — “So bottom line: if you’re reaching for something after work, the GDP is your move”

​
Cost Estimates

ComponentCost Per EpisodeNotes
Script generation (LLM)0.02−0.02 - 0.08~2,000 tokens in, ~1,500 tokens out
TTS Host A (2.5 min)0.04−0.04 - 0.15ElevenLabs pricing varies by plan
TTS Host B (2.5 min)0.04−0.04 - 0.15Same voice API, different voice
Audio processing0.01−0.01 - 0.02FFmpeg, minimal compute
CDN storage (5MB MP3)$0.001Negligible per file
Total per episode0.10−0.10 - 0.40Lowest cost of any AI media format

​
Tier Impact

TierAccess
Free15-second faded preview of each episode. Audio plays for 15 seconds with the hosts mentioning the user’s name and strains, then fades to silence with an upgrade prompt. Enough to hear the personalization, not enough to get the value.
ProUnlimited episode generation for all formats. Full playback, download as MP3, share clips. Auto-generation for every report and monthly recap.
StudioCustom host voice selection (choose from 6+ voice pairs), extended Monthly Recap episodes (15-20 min), “Season” compilation of all monthly recaps, priority generation queue.

​
Future Possibilities

  • Choose host personalities — Casual, scientific, funny, concise. Let users pick the vibe.
  • Friend collaboration episodes — Two users opt in to a combined episode comparing their cannabis preferences: “John prefers indicas, but Sarah is all about the sativa hybrids…”
  • Seasonal specials — A “420 Special” recap episode, “Year in Review” annual episode, “New Year’s Strain Predictions” episode.
  • Guest experts — A third AI voice representing a “cannabis scientist” or “budtender” for specific topics.
  • Podcast feed — Generate a private RSS feed so users can listen in their podcast app (Apple Podcasts, Spotify) alongside their regular shows.

​
Dependencies

  • Research report generation pipeline — built and live
  • User order history and stash data — built and live
  • Trigger.dev task infrastructure — built and live
  • Blog article content — built and live
  • Dialogue script generation prompt (LLM)
  • TTS integration with dual-voice support (ElevenLabs)
  • Audio assembly pipeline (FFmpeg or similar)
  • In-app audio player component with scrubbing
  • Push notification for “episode ready” event
  • CDN storage for audio files
  • 15-second preview extraction for upgrade hook

​
Open Questions

  1. NotebookLM API vs. custom pipeline — Google may offer a NotebookLM API that handles the entire dialogue-to-audio pipeline. If available, it could dramatically simplify implementation. But it would also mean depending on Google’s API availability, pricing, and quality. Custom gives us full control. Monitor NotebookLM API announcements.
  2. Episode length sweet spot — NotebookLM episodes run 10-20 minutes, which works for dense documents. Our data is more focused (one order, one strain). Is 3-5 minutes the right length, or do users want longer episodes? Need user testing.
  3. Background generation timing — Should podcast generation start immediately when a report is generated (user might not care about podcasts) or wait until the user visits the report page (saves cost but adds latency)? Recommendation: auto-generate for the first 3 reports to demonstrate value, then switch to on-demand if the user has not listened.
  4. Sharing format — Share the full episode MP3, a 30-second clip, or a visual audiogram (waveform animation with subtitles over a branded background)? Audiograms perform best on social media but add rendering complexity.
  5. Voice consistency — Should the hosts always sound the same (familiar, ritual-building) or should the voice profiles vary slightly to prevent listener fatigue? NotebookLM uses consistent voices, which seems to work.