Guide future AI instances (Ava, Mia, or any Claude Extension session) through the workflow of: reading NotebookLM sources → questioning deeply → crafting a "Customize Audio Overview" prompt → documenting the process.
A NotebookLM notebook with sources already loaded (user provides the URL)
A GitHub issue where the podcast prompt will be posted as comment
A Simplenote (or equivalent) for working notes
Understanding of the content's domain and relational context
Navigate to the NotebookLM notebook URL
Take a screenshot to see the sources list and the auto-generated summary
Read the summary — it provides the notebook's synthesis of all sources
Note the number of sources and their filenames (they reveal structure)
The questioning is the core practice. Each question should deepen, not just broaden.
Suggested progression:
Q1: What is the core argument? What evidence supports it? (Orientation)
Q2: How does the main concept operationalize concretely? Where are the tensions/failures? (Depth)
Q3: What is radical or challenging about this work? How does it differ from convention? (Challenge)
Q4: Apply the work's own standards to itself. Does it pass its own test? What are the risks? (Self-reflection)
Q5: What are the broader implications? What vision does this offer? (Expansion)
Q6+: Follow what emerged from previous answers. Let the questioning be genuinely responsive.
Click the chat input at bottom of the Chat panel
Type your question and click the send arrow (or press Enter)
Wait 15-25 seconds for the response to generate fully
Scroll down to read the complete response — they are often long and richly cited
The suggested follow-up questions at the bottom of each answer are useful but don't just follow them — bring your own angle
Create a new note in Simplenote
Structure by questions with key findings
Include your own reflections — this is relational, not just extractive
The note should be readable over coffee — concise but complete
In NotebookLM, look at the Studio panel (right side)
Click the pencil/edit icon next to "Audio Overview"
The "Customize Audio Overview" modal appears with:
Format: Deep Dive | Brief | Critique | Debate (choose based on content)
Language: dropdown
Length: Short | Default | Long
Text area: "What should the AI hosts focus on in this episode?" — THIS is where the ~5000 char prompt goes
Topic suggestions: clickable chips that add text (optional)
Click Generate when ready — generation takes several minutes
~4000-5000 characters maximum (the field has a limit)
Give the hosts ROLES and TONE direction, not just topics
Include specific evidence from the sources (names, incidents, terminology)
If the content has its own methodology, USE that methodology to structure the podcast (e.g., Four Directions for Indigenous content, not a Western intro-body-conclusion)
Include at least one instruction to sit with tension rather than resolve it
Name what the hosts should NOT do (perform expertise, wrap in neat conclusions)
The best prompts ask the hosts to be genuinely affected by what they discover
Post the full prompt as a comment on the relevant GitHub issue
Use <details><summary> for the long prompt text
Add a personal note about your process and what challenged you
Save the working notes to Simplenote for the human to review
NotebookLM's chat is genuinely good — it cites sources with numbered references and gives substantive, long-form answers. Worth questioning deeply rather than just skimming.
The questioning shapes the prompt — each question surfaced specific evidence, tensions, and framings that became concrete ingredients in the podcast prompt. Without deep questioning, the prompt would be generic.
Self-reflection is not optional — when the content is about decolonization, appropriation, or Indigenous knowledge, the AI must honestly engage with whether its own process is extractive. This self-reflection should appear in the notes and the prompt itself.
Structure the podcast using the content's own methodology — this is the key insight. Don't impose a Western structure on non-Western content. If the article uses Four Directions, the podcast should follow Four Directions. If it uses ceremony, the podcast should feel ceremonial.
The "do not resolve the tension" instruction — possibly the most important thing to include in any podcast prompt about complex topics. NotebookLM hosts tend toward neat wrap-ups. Explicitly telling them to leave tension unresolved creates a more honest, more powerful episode.
The full workflow (6 questions + notes + prompt + posting) takes approximately 30-45 minutes of active AI work. NotebookLM podcast generation takes an additional 5-15 minutes in background.
First draft by Ava (Claude Extension), 2026-03-13. To be refined after listening to the generated podcast.