Creator publishes 20-prompt NotebookLM research system for 500-page source packs

• Heyrimsha’s thread opener packages NotebookLM as a repeatable research workflow, claiming it can turn a 500 page source pack into clear answers in under an hour.
• The strongest early prompts in the Source Onboarding prompt, the Contradiction Hunter, and the Blind Spot Detector all target the same job: map the corpus before you start writing.
• The Evidence Auditor and the Insight Extractor push NotebookLM past generic summarization and toward claim checking, novelty hunting, and source ranking.
• Google’s own NotebookLM product page and help docs back the basic mechanics here: upload up to 50 sources, chat against them with inline citations, and generate artifacts like briefings, study guides, and mind maps.

You can trace the workflow from the public thread to Google’s own source upload docs, which say each notebook can hold up to 50 sources and each source can run as large as 500,000 words or 200MB. The interesting part is how closely the thread lines up with official NotebookLM outputs: the chat guide emphasizes source-grounded answers with hoverable citations, while the notebook creation docs list reports, FAQ sheets, briefing docs, study guides, audio overviews, and exports to Docs or Sheets. The twist is that none of this is a product launch. It is one creator’s prompt stack, published in public, sitting on top of an existing research tool.

The first prompt is the cleanest one in the set. It asks NotebookLM for four things before any real questioning starts: the top themes, where sources agree or clash, the most surprising finding, and the biggest open questions.

That matches NotebookLM’s official pitch unusually well. The product page describes the app as a research partner that summarizes uploaded PDFs, websites, videos, audio, Docs, and Slides, then makes connections across them. The source docs also confirm that notebooks are built around static source copies, so the value here comes from forcing a map of the corpus up front.

The next cluster is built for adversarial reading, which is where this thread gets more useful than a normal prompt dump.

• The Contradiction Hunter asks for conflicting claims, named sources, and an evidence-weighted judgment about which side is stronger.
• The Blind Spot Detector asks what is missing, which voices are absent, and which shared assumptions nobody questions.
• The Evidence Auditor adds a confidence layer by rating claims as weak, moderate, or strong and separating anecdotal, correlational, and causal evidence.

Google’s chat documentation says NotebookLM can answer questions and perform actions using direct quotes, text, and images from uploaded sources as citations. That makes these prompts less about magic phrasing and more about giving the model a hard editorial brief.

The middle of the thread shifts from analysis to packaging. The Executive Brief Builder asks for a 250 word brief with one lead finding, three evidence-backed points, one uncertainty, and a closing implication.

Several later prompts extend the same pattern:

• The Timeline Builder turns a pile of documents into shifts in consensus over time.
• The Analogy Engine reframes complex concepts for a general audience, while forcing the model to name where each analogy breaks.
• The Follow-Up Question Generator asks for the next ten unanswered questions and the source types that might answer them.

That maps directly onto NotebookLM’s official output surface. The notebook creation guide says the Studio panel can generate reports, FAQ sheets, study guides, briefing documents, and exports to Docs or Sheets, so a creator working on a script, explainer, or treatment could use these prompts as scaffolding rather than starting from a blank chat box.

The last stretch adds a useful sequencing rule. The thread’s 19th and 20th prompts are explicitly saved for the end: one builds first, second, and third order implications while flagging where speculation begins; the other asks for the single new thing learned, the safest claim to cite publicly, the area that still needs evidence, and a three sentence summary of the whole project.

That ending is new information, not recap, because it turns the thread into a closed loop: map the sources, stress test the claims, generate outputs, then force a final declaration of what is known and what still is not. For NotebookLM users who already rely on source-grounded chat and generated artifacts, the thread is basically a ready-made operating manual.