Generating a podcast

Aparture doesn't produce audio directly. Instead, it builds a bundle of markdown sources plus a focus prompt that you upload to NotebookLM — Google's hosted tool — which generates a two-host conversation about the day's papers. The whole flow takes a few minutes of active clicking on top of whatever NotebookLM needs to render the audio.

This page covers what's in the bundle, how the upload works, and the two places where you can tune podcast style.

Is this worth setting up?

The podcast flow is entirely optional. Plenty of users skip it and just read their briefings. It tends to earn its keep if you commute or exercise regularly and want to keep up passively, if audio sticks better for you than text, or if you want to share a briefing with a collaborator who isn't in Aparture. If most of your paper reading happens at a desk and you don't want another tool in the loop, skipping this add-on is perfectly reasonable.

What's in the bundle

Once a briefing exists, a NotebookLM Podcast card appears below it in the main area. Clicking Generate NotebookLM bundle downloads a ZIP file containing:

File	Purpose
INSTRUCTIONS.md	The three-step upload workflow you're about to do. Not a NotebookLM source — read it and keep it local.
briefing.md	The full briefing rendered as a single markdown source. Executive summary, themes, paper cards.
discussion-guide.md	A podcast outline generated from the briefing — theme weighting, talking points, pacing, scaled to duration. The main "content" source NotebookLM reads from.
papers/<arxivId>.md	One file per paper, holding the Stage 4 deep-analysis report. NotebookLM uses these for detail during the conversation.
focus-prompt.txt	The prompt you paste into NotebookLM's audio-customisation box. Tells the hosts what to emphasise, how deep to go, and how to pace the conversation.

Everything is plain markdown or text, so you can inspect the bundle before uploading.

Generating the bundle

Inside the NotebookLM card:

1. Duration. Drop-down with 5, 10, 15, 20, and 30 minute options. The discussion guide scales to the chosen length — a five-minute episode covers one theme and 2–3 papers at headline depth, a thirty-minute episode covers 3–4 themes with meaningful dives on the top papers. The scaling rules live in lib/notebooklm/buildFocusPrompt.js if you want to see the exact budget per tier.
2. Model. Drop-down for which model generates the discussion guide. Picks up every model in AVAILABLE_MODELS — pick a capable one, since this is a synthesis task. claude-sonnet-4.6 or gemini-3.1-pro are reasonable defaults.
3. Generate. Click Generate NotebookLM bundle. Takes roughly 30–60 seconds; progress shows inline.

When it's done, a ZIP file downloads with a name like aparture-notebooklm-2026-04-17.zip.

Uploading to NotebookLM

NotebookLM doesn't offer an API for notebook creation or audio generation, so this is a manual upload flow. Expect 2–3 minutes of clicking. Step-by-step (the version in INSTRUCTIONS.md inside the bundle stays current if Google changes the UI):

1. Unzip the bundle locally. NotebookLM doesn't accept ZIPs as sources.
2. Open notebooklm.google.com and sign in with a Google account if needed.
3. Create a new notebook.
4. In the Sources panel, choose "Add source" → "Upload files" and select every .md file from the unzipped bundle at once — briefing.md, discussion-guide.md, and all of papers/<arxivId>.md.
5. Do not upload focus-prompt.txt or INSTRUCTIONS.md as sources. They're for the next step and for you, respectively.
6. Wait a few seconds for NotebookLM to index the sources.
7. In the main panel, find the Audio Overview controls and click Customize (sometimes a pencil icon).
8. Paste the entire contents of focus-prompt.txt into the customisation text box.
9. Click Generate. NotebookLM takes a few minutes to render the audio — timing varies with queue depth on Google's side.

When it finishes, you'll have a 5–30 minute audio overview in the notebook. Play it in-browser, download it, or share the notebook link.

NotebookLM UI changes

Google revises the NotebookLM interface periodically. Button names and panel locations drift. The bundle's INSTRUCTIONS.md is the version of these steps that ships with your bundle; if something doesn't match, trust that one over this page.

NotebookLM ignores the target duration

The Audio Overview setting (Shorter / Default / Longer) is NotebookLM's own pacing knob, and it overrides the duration target in the focus prompt more often than not. Expect:

• Default: typically 60–80% of the target (a 30-min prompt often yields ~20 min).
• Longer: typically 130–170% of the target (a 30-min prompt often yields ~45 min).

Pick the setting that's closer to what you actually want, then accept the ±30% variance. The focus prompt asks the hosts to prioritise keeping the number of papers and depth intact when length is stretched or compressed, rather than dropping content.

What the focus prompt does

The focus prompt is what distinguishes a generic "read the sources aloud" podcast from one that sounds like two researchers talking through the day's papers. It carries:

• Your research context, paraphrased from the profile text.
• How many papers to cover and at what depth, scaled to the duration you picked.
• What kind of conversation to have — argument-focused, willing to surface tensions, careful with citations.
• What to skip, usually purely descriptive or low-scored material.

The focus prompt is generated when you build the bundle, so it already reflects the specific briefing and duration. You don't need to edit it — paste it in as-is.

Tuning podcast style

Two places shape the output, and they behave differently.

prompts/notebooklm-discussion-guide.md — the outline template

This is the prompt that produces discussion-guide.md. It controls how themes are ordered and weighted in the outline, which papers get deep-dive coverage versus brief mentions, the overall conversation arc (open, themes, transitions, close), and the pruning rules when the outline would overrun the duration budget.

Edit this file and the next bundle you generate will use the new template — the server re-reads it on each call, no rebuild needed. This is usually the right place to experiment with podcast structure.

lib/notebooklm/buildFocusPrompt.js — the focus prompt builder

This is a JavaScript function (not a markdown template) that produces focus-prompt.txt. It controls the duration-scaled depth strategy — how time is allocated across deep-dives versus brief mentions at each duration tier — along with theme-count scaling and the conversation-style directives embedded in the focus prompt.

Because this lives in JS, changes require a server restart to take effect. It's less convenient for quick iteration but more powerful, since you can condition the prompt on briefing content or duration programmatically.

For the full prompt reference, see Reference: prompts.

Common issues

The audio sounds generic. Usually the focus prompt wasn't pasted, or it was pasted into the wrong text box. The Customize → text-box flow is what shapes the podcast; without it, NotebookLM produces a default-tone summary.

The hosts skip a paper you wanted covered. At shorter durations, the outline prunes aggressively by score. Star the paper in the briefing, re-generate the bundle, and re-upload. Starred papers get priority in the discussion guide.

Length is way off target. NotebookLM's Audio Overview setting controls pacing more than the focus prompt does — see the warning box in the upload section. Pick Shorter / Default / Longer to get closer to your target and accept ±30% variance.

The hosts spend the first minute or two on preamble and banter. The focus prompt explicitly tells the hosts to skip greetings, agenda-setting, and meta-narration about the format, but NotebookLM sometimes inserts them anyway. The longer the Audio Overview setting, the worse this tends to be. Shorter tier has less runway for filler.

The hosts repeatedly announce what they'll do instead of doing it ("we'll cite arXiv IDs today", "we'll keep this conversational", etc.). Same root cause as banter — NotebookLM sometimes meta-narrates despite the focus prompt telling it not to. A shorter Audio Overview setting usually cuts this down because there's less time to fill.

NotebookLM hits a context limit or truncates. Too many sources can do this. Try uploading only briefing.md and discussion-guide.md, dropping the per-paper reports. NotebookLM can usually infer enough from the briefing alone.

Bundle generation takes a long time. The LLM call for the discussion guide is the slow part. Switching notebookLMModel to a faster model (Gemini 3 Flash, Claude Sonnet 4.6) cuts this noticeably.

Next

• Reading a briefing — the briefing is the upstream input, so tuning it tunes the podcast.
• Tuning the pipeline — thresholds and model choices that shape what reaches the bundle.
• Reference: prompts — every prompt file Aparture ships, including the discussion-guide template.