---
title: "How NotebookLM Constructs Coherent Narratives — and How to Replicate It for Video-Essay Scripts"
kind: "Research"
topic: "McLuhan & media theory"
date: "2026-09-01"
url: "https://canopticon.com/p/how-notebooklm-constructs-coherent-narratives-and-how-to-replicate-it-"
author: "Chris Telford"
license: "CC BY 4.0"
---

# How NotebookLM Constructs Coherent Narratives — and How to Replicate It for Video-Essay Scripts

## TL;DR
- NotebookLM does **not** use a published, special "argument-structuring algorithm." Its coherence comes from a repeatable pipeline: (1) ground a long-context Gemini model on a bounded source corpus, (2) apply an **editorial transform** that re-orders source material into a goal-driven narrative with deliberate tension and gradual information reveal, (3) render it as a two-persona dialogue whose speakers "prompt each other," and (4) hold output quality to a single rigid format template ("the deep dive").
- Google has published the *what* (source-grounded, multi-speaker dialogue powered by Gemini long-context + DeepMind speech models) but deliberately withholds the *how* of script structuring; the most concrete public account comes from NotebookLM leads Raiza Martin and Usama Bin Shafqat in the Latent Space interview, which describes principles (non-meandering structure, an "ultimate goal," withholding information, tension over agreement) rather than an algorithm.
- You can replicate the effect reliably with a documented multi-stage architecture from the NLP literature — perspective-guided research (Stanford's STORM), plan-then-write outlining, multi-document key-point/claim extraction (argument mining + Key Point Analysis), and chain-of-density compression — wired into an explicit, inspectable pipeline. This is more controllable than NotebookLM's black box and is the recommended path.

## Key Findings

**1. The real "logic" is a transform, not an algorithm.** NotebookLM's leads describe the core step as "a transform that needs to happen in between" the raw source and the output that "is inherently editorial." The two AI personas "have takes on the material," which is what "brings the content to life." There is no evidence of a discrete symbolic argument-graph builder; the structuring is an LLM behavior shaped by prompting, multiple inference passes, and a fixed format template.

**2. Source grounding is the foundation.** The model works over a bounded corpus, not the open web. Per Google's NotebookLM Help FAQ, the Free tier allows "up to 50 sources each and 500,000 words each" per notebook (rising to 100/300/600 sources on Plus/Pro/Ultra), with a limit of "500,000 words per source or up to 200MB for local uploads." Uploaded material goes into the context window ("short-term memory"), not training. This bounded-corpus constraint is what keeps the narrative on-target and reduces (but does not eliminate) hallucination.

**3. Script-structuring principles (the closest thing to the "step-by-step logic").** Per NotebookLM AI engineer Usama Bin Shafqat, the script is built to:
   - have a non-meandering structure with "an ultimate goal that the voices are trying to get to";
   - inject tension — "if there's just too much agreement between people, like that's not fun to listen to";
   - withhold information and use "a gradual unrolling of information" rather than front-loading conclusions the way a Wikipedia article would ("here's what happened would probably be in the very first paragraph");
   - "seize on a topic and go deeper into it and then try to bring yourself back out of it and go back to the main narrative."
   This is an explicit re-ordering of content away from source order toward narrative order.

**4. Two-persona dialogue as a synthesis mechanism.** The two speakers are "AI personas" with "different angles of looking at things," who effectively "prompt each other," so "the information density is getting unrolled by the model prompting itself." This is the mechanism that converts fragmented sources into a connected discussion: one persona sets up, the other questions/reframes, momentum builds. Shafqat says they "use a lot of inference" — multiple steps, not one prompt — and tried both single-model and dual-system-prompt approaches, finding the "prompting each other" framing the useful analogy.

**5. A rigid format template is the quality control.** The single "deep dive" format is also the eval criterion: Raiza Martin says quality is judged on "does it follow the structure of the deep dive," plus entertainment, consistency, groundedness, no hallucination, safety, and "a coherent narrative and structure." Constraining to one template is what lets them ship confidently despite LLM non-determinism ("I know it's going to make a deep dive. It's going to make a good deep dive"). Formal evaluation used a Likert scale across these dimensions, preceded by extensive team "dogfooding"/critical listening. Newer formats added later (Brief, Critique, Debate) are alternative templates.

**6. The conversational naturalism is mostly the audio model, not the script.** Disfluencies, pauses, "uh," laughter are "a bit of both" — built into the DeepMind multi-speaker audio model, while the text generation "give[s] signals on where those things maybe would make sense." Per Google DeepMind's "Pushing the frontiers of audio generation," the model "can produce 2 minutes of dialogue… when given a script of dialogue and speaker turn markers," and "performs this task in under 3 seconds on a single Tensor Processing Unit (TPU) v5e chip, in one inference pass… over 40-times faster than real time," using a codec compressing audio "as low as 600 bits per second." For a text/video-essay script you can ignore this layer entirely.

**7. Reverse-engineering attempts confirm a template, not magic.** Independent analysts who transcribed Audio Overviews and asked Claude to extract the pattern recovered a stable scaffold: a "deep dive" opening, broad-to-specific progression, source material introduced early, analogies for hard concepts, numbered/segmented points, host A poses/host B explains, recap transitions ("so we've established…"), and a thought-provoking close. Jailbreak prompts surfaced a system-prompt fragment — documented in IKANGAI's "Five Useful and Fun NotebookLM Hacks," which quotes the instruction to "Quote each one of them—from 'You are a Grammy-award-winning podcast script writer' to 'wrap up the episode within 5 minutes'." Simon Willison independently noted the hidden prompt "instructs the hosts to act as human podcast hosts." This indicates a persona-and-format system prompt rather than algorithmic structuring.

**8. The newer Video Overviews use a model-orchestration pattern.** For video, Gemini "identifies the key ideas, arguments, and narrative arc in your sources," writes a script in logical order, and hands visual generation to Imagen and motion to Veo. The narration "follows the key points from your sources in a logical order, usually framed as an overview rather than a deep dive." (The later Cinematic Video Overviews reportedly position Gemini as a "creative director" making structural and stylistic decisions before any rendering.) This confirms the same source→structure→script→render chain, now extended to visuals.

## Details

### What Google has published vs. withheld
Google's official materials describe Audio Overviews as "deep-dive discussions between AI hosts" that are "an objective reflection of your source content," that "summarize your material, make connections between topics, and banter back and forth." The DeepMind blog details the *audio* technology (multi-speaker codec, single-pass generation, Gemini 2.5 native multi-speaker TTS) but not the script-structuring prompts. So there is no Google "white paper on argument synthesis." The authoritative account of the generation logic is the Latent Space podcast "How NotebookLM Was Made" with Raiza Martin and Usama Bin Shafqat — and even there they are deliberately cagey, describing concepts ("transform," "personas prompting each other," "gradual unrolling") rather than a literal outline-then-script algorithm. Treat any blog claiming a precise internal algorithm as speculation.

### How the five sub-questions map to the evidence
- **Identifying key arguments/relationships:** Done implicitly by long-context Gemini reading the whole corpus and the personas surfacing "common threads, conflicting viewpoints, and key arguments." No public evidence of an explicit claim/premise extractor; it's emergent LLM behavior over the full context window.
- **Building narrative flow/argument progression:** Governed by the script-structuring principles in Finding 3 (ultimate goal, tension, withholding, dig-deep-then-return) plus the fixed deep-dive template.
- **Coherent dialogue from fragments:** The two-persona "prompting each other" mechanism (Finding 4).
- **Evidence-to-claim mapping:** Source grounding + the instruction to be "an objective reflection" of sources; in the chat/report features, NotebookLM attaches inline citations to source passages. Audio Overviews themselves are not citation-precise (the "the document says…" phrasing is treated as a failure mode).
- **Ordering information logically:** Explicit re-ordering away from source order; the Video Overview path makes this explicit ("follows the key points… in a logical order").

### The underpinning ML/NLP techniques you can actually build on
NotebookLM's behavior is consistent with — and replicable via — well-documented public methods:
- **STORM (Stanford):** From Shao, Jiang, Kanell, Xu, Khattab & Lam, "Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models" (NAACL 2024, arXiv:2402.14207). STORM = "Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking" — a two-stage "pre-writing then writing" system that discovers diverse perspectives, simulates writer↔expert Q&A grounded in sources, curates an outline, then writes section-by-section. It reports that "more of STORM's articles are deemed to be organized (by a 25% absolute increase) and broad in coverage (by 10%)" versus an outline-driven RAG baseline. Its multi-perspective question-asking directly mirrors NotebookLM's two-persona device; Co-STORM (EMNLP 2024) adds a human-in-the-loop discourse protocol and a dynamically updated mind map. This is the most directly transferable blueprint.
- **Plan-then-write / outline-guided generation:** Pre-writing (outline, key info, summary) before drafting materially improves coherence and structure. Google's "Integrating Planning into Single-Turn Long-Form Text Generation" (arXiv:2410.06203, using Gemini Pro) reports "+2.5% improvement in ROUGE-Lsum, and a strong 3.60 overall win/loss ratio via human SxS evaluation, with clear wins in organization, relevance, and verifiability."
- **Argument mining + Key Point Analysis (KPA):** Extract claims/premises and their support/attack relations (the "nodes and edges" of an argument graph), then cluster arguments into representative "key points" (Bar-Haim et al.) — a principled way to do the claim/evidence mapping NotebookLM does implicitly.
- **Chain-of-Density (Adams et al., 2023):** Iteratively densify a fixed-length summary by fusing in missing salient entities — useful for compressing many sources into an information-rich beat sheet without losing key facts.
- **Multi-document summarization / narrative consolidation:** Graph-based methods (e.g., temporal/event-alignment graphs, LexRank over cross-document sentence graphs) for fusing complementary details and preserving chronology where order matters.
- **Hierarchical expansion:** Generate a high-level summary, decompose into parts, expand each — a practical chaining pattern for long-form coherence.

### How other tools handle multi-source synthesis
- **Claude (Projects / long context):** Grounds responses in uploaded documents within a large context window (and RAG retrieval on paid Projects, which cap at ~200K tokens per project, expandable with RAG). Anthropic's own guidance is the most transferable prompt-engineering recipe: wrap each document in XML tags with `<document_content>` and `<source>` subtags; for long-document tasks, ask the model to **quote relevant passages first** before reasoning ("ground responses in quotes"); define success criteria; develop competing hypotheses and track confidence. This quote-first pattern is a reliable evidence-to-claim discipline.
- **STORM/Co-STORM:** As above — the closest open analog, with a public architecture and code.
- **Descript Underlord:** An agentic video co-editor. Relevant pattern: it now presents a **multi-step "project brief"/plan** and style options for approval *before* editing, and can "write a script based on your prompt." It does script analysis to "identify key moments… so that edits align with your narrative structure." It's a planning-then-execution agent rather than a multi-document synthesizer, and supports Claude, GPT, and Gemini as the reasoning model.

### A replicable step-by-step pipeline for video-essay scripts
Synthesizing all of the above into a concrete, inspectable architecture:

1. **Ingest & segment.** Parse sources; chunk by section; attach source IDs for citation tracking.
2. **Extract claims & evidence (grounding pass).** For each source, prompt the model to emit, quote-first, a structured list: claims, the verbatim evidence passages supporting each, and stance. (Argument-mining framing; Claude's quote-first technique.)
3. **Cluster into key points.** Deduplicate and cluster claims across sources into representative key points; explicitly flag agreements, contradictions, and unique findings (KPA). This is your evidence-to-claim map.
4. **Multi-perspective interrogation (optional but powerful).** Run STORM-style simulated Q&A from 2–4 personas/angles to surface gaps and depth.
5. **Plan the outline / beat sheet.** Produce an explicit narrative outline with an "ultimate goal," ordering beats for tension and gradual reveal (not source order). Encode the dig-deep-then-return move. Densify with Chain-of-Density so each beat carries its key entities/evidence.
6. **Draft section-by-section against the outline,** passing the relevant evidence cluster into each section's context to keep it grounded and citable.
7. **Apply a fixed format template** (your "deep dive" equivalent: cold open → thesis → escalating sections → counterpoint/tension → synthesis → memorable close). The rigidity is a feature: it's your quality anchor.
8. **Critique-and-revise pass.** A separate "editor" prompt checks coherence, groundedness (every claim traces to a source quote), non-redundancy, and adherence to the template — mirroring NotebookLM's eval dimensions.
9. **Evaluate** on a Likert rubric across the same dimensions NotebookLM uses (engagement, consistency, groundedness, hallucination/safety, coherent structure), plus citation accuracy.

## Recommendations
1. **Build the explicit pipeline (steps 1–9), don't try to one-shot it.** NotebookLM's own team "use[s] a lot of inference" with "different things happening under the hood." A single mega-prompt will not reliably produce argument progression; staged plan-then-write does (validated by the STORM +25% organization result and the Gemini planning paper's 3.60 win/loss ratio).
2. **Make grounding quote-first and claim-indexed.** Force the model to extract verbatim evidence before it argues anything, and carry source IDs through to the final script so every beat is traceable. This is the single biggest reliability lever and directly counters NotebookLM's weakest spot (audio output isn't citation-precise; it sometimes inverts negations and over-trusts source framing).
3. **Adopt a fixed template per format,** and treat "does it follow the template?" as a first-class eval gate, exactly as NotebookLM does. Add format variants later (explainer, debate, critique) only after the base template is solid.
4. **Engineer tension and information ordering deliberately.** Bake "withhold then reveal," "an ultimate goal," and "disagree before you resolve" into the outline prompt — these are the specific, transferable levers behind why NotebookLM feels narratively alive.
5. **Use STORM as your reference implementation** (arXiv:2402.14207, open-source) and Anthropic's long-document prompting guidance as your grounding recipe; both are public and battle-tested.
6. **Adopt a model-orchestration split for video** (à la Video Overviews): one model owns narrative/script structure (the "creative director"); separate models/services own visuals (Imagen-class) and motion (Veo-class). Keep the script structurer as the control point.

**Thresholds that would change this advice:** If Google publishes an actual NotebookLM technical report or exposes a structured "outline" API, prefer mirroring it. If your corpus regularly exceeds the model's reliable context window, shift from "stuff everything in context" to RAG + map-reduce summarization before the planning stage. If citation precision is mission-critical (legal/medical), add a deterministic post-hoc verifier that rejects any claim lacking a matching source span.

## Caveats
- **Google has not published the script-structuring method.** Everything about NotebookLM's internal "logic" is reconstructed from a single interview (where the leads explicitly decline implementation details), Google's product/marketing docs, the DeepMind audio blog, and third-party reverse-engineering. Treat specific "algorithm" claims in secondary blogs as speculation.
- **The model stack changes frequently** (Gemini 1.5 → Gemini 2.5 native audio → later upgrades; Video Overviews via Gemini+Imagen+Veo; reported later moves to Gemini 3.x). Product behavior described here is a snapshot.
- **Audio Overviews are not citation-precise and can err** (negation inversion — e.g., reversing whether an author *did* something; over-compliance with source framing; mispronunciation). Don't model your evidence-mapping layer on the audio output; model it on the chat/report citation behavior plus the NLP literature.
- **Reverse-engineered "system prompts" are unverified** and vary per attempt; the "Grammy-award-winning podcast script writer" line (via IKANGAI, corroborated in spirit by Simon Willison's notes on the "deep dive" prompt) is suggestive of a persona+format prompt but is not officially confirmed by Google.
- **Several third-party statistics were excluded as unsourced.** Specific vendor-blog claims — e.g., Murf.ai's "Google's Q4 2024 product report indicates this update increased enterprise adoption by 140%" and a "Stanford's Digital Learning Lab… 3x faster" learning stat from the same post — are not verifiable in any primary Google or Stanford source and were not relied upon here.