Write a Whitepaper or Ebook with Claude + NotebookLM: The Workflow I Actually Use

Two months ago I shipped a 34-page whitepaper for a B2B martech client. Topic was "first-party data strategies for SaaS marketers after cookie deprecation." Research, outline, draft, fact-check, design brief, landing page copy — the full pipeline, 2.5 days from kickoff to gated PDF live. Three years ago, the same scope would have taken three weeks and at least one junior strategist's sanity.

The combo that did most of the heavy lifting: Claude Opus 4.6 for the writing, structuring, and judgment work, and NotebookLM for the research intake, source-grounded fact-checking, and the citation engine that makes the final document defensible. They overlap in some places and complement in others. Used wrong, they fight each other. Used right, they cut lead-magnet (获客内容) production to a small fraction of what it used to cost.

Here's the exact 6-stage workflow, the prompts, and the places where I had to learn the hard way.

Why the Combo Works (and Why Most People Get It Backwards)

The first mistake I see: marketers use one tool for the whole job. They either try to make NotebookLM write the whitepaper (it won't — it's a research surface, not a writer), or they make Claude do the research from its general knowledge (it will — but you'll get confident, unsourceable claims that fall apart under client scrutiny).

The right division:

• NotebookLM = your research library and fact-checker. It ingests source material, indexes it, and answers questions with citations to the exact paragraph. It also generates audio summaries, mind maps, and briefing docs from the same library. It does not invent — if a claim isn't in the uploaded sources, it tells you so.
• Claude Opus 4.6 (or Sonnet 4.6 for budget runs) = your writer, editor, and structural thinker. It reasons across long context, holds the thread of a 30-page document, and produces prose at a level ChatGPT still can't match for serious B2B (Business-to-Business, 企业对企业) work.

The mistake is mixing them up. Don't ask Claude to "go research this topic" — it will hallucinate. Don't ask NotebookLM to "write the whitepaper" — it will refuse or produce a dry summary. Use each for what it's actually good at, and the workflow moves at 5–10x.

The 6-Stage Workflow

Stage 1 — Lock the Topic and the Angle (30 minutes)

Don't start with "write a whitepaper about AI in marketing." Start with a buyer problem your client (or your business) is already getting asked about. The whitepaper is supposed to do two jobs: rank for a high-intent query and convert that traffic into leads. The topic and angle should be decided before you open any AI tool.

What I need locked before stage 2:

1. One specific reader persona — title, company size, the exact problem they're trying to solve.
2. One specific promise — "By the end of this document, you'll know how to [concrete outcome]."
3. A short list of 5–10 source documents I already trust on the topic — analyst reports, original research, internal data, customer interviews, prior posts that performed well.

For the martech whitepaper, the promise was "a defensible 90-day roadmap for collecting, organizing, and activating first-party data without third-party cookies." Sources were 4 analyst reports, 2 internal customer interview decks, and 3 competitor whitepapers I'd already pulled for SEO work.

Stage 2 — Research Library in NotebookLM (1–2 hours)

Drop every source document into a single NotebookLM notebook. PDFs (Portable Document Format, 可携带文档格式), Google Docs, plain text, even YouTube transcripts — it indexes them all.

Critical detail: upload more than you think you need. The cross-referencing between documents is where NotebookLM beats manual note-taking. If I upload the analyst reports, internal interviews, and competitor whitepapers all in one notebook, NotebookLM can answer a question like "what do these 9 sources say about email deliverability as a first-party data source?" with citations from each.

What I do at the end of this stage:

• Ask NotebookLM to summarize each source in 5 bullets with a confidence flag.
• Generate a briefing document in the "Report" output format — it's the cleanest research briefing I get from any tool, and it becomes the working brief I feed into Claude.
• Build a "facts" notebook alongside — a separate notebook where I dump the 30–50 specific facts, stats, and quotes I want the whitepaper to use. This becomes the citation backbone in Stage 5.

NotebookLM's "Audio Overview" feature is a sleeper hit for this stage. The 10-minute podcast-style summary of the source library is surprisingly good at surfacing connections I missed. I listen to it on a walk and flag anything worth pulling into the brief.

Stage 3 — Outline in Claude (45 minutes)

Now Claude enters. I feed it the NotebookLM briefing doc plus the persona and promise from Stage 1, and ask for a section-by-section outline.

The prompt I use:

You are a senior B2B content strategist. I'm writing a 30-page lead-generation whitepaper on [TOPIC] for [PERSONA]. The goal: rank for [TARGET QUERY] and convert readers into sales conversations. Attached is a research briefing pulled from 9 primary sources (analyst reports, customer interviews, competitor assets). All claims in the final draft must be traceable to the briefing.

Produce a section-by-section outline with:

1. Working title and 3 alternative titles tested for SEO difficulty
2. 6–8 main sections, each with a 1-sentence thesis
3. For each section: 3–5 specific sub-points, each tied to a fact or quote from the briefing (cite source)
4. A logical reading flow — what does the reader need to know in section 3 to make section 4 land?
5. Where the data visualizations should go
6. A short executive summary (200 words) the reader can absorb in 90 seconds

Push back on my angle if you see a stronger one in the sources.

That last line is important. Claude Opus is genuinely good at saying "your angle is weaker than this one I see in the data." I take that feedback roughly 60% of the time.

Stage 4 — Drafting in Claude (3–5 hours, broken into chunks)

This is where most AI whitepaper attempts die. The model produces 2,000 words of generic thought-leadership mush, and the writer accepts it because it's "done."

Three rules that fixed this for me:

Rule 1: Draft section by section, never the whole document at once. A 30-page whitepaper is too long for any current model to hold perfectly in one pass. I ask for 800–1,200 words per section, in order, with a 200-word recap of all prior sections fed in each time. This is where Claude's long context window matters — Opus 4.6 holds 1M tokens, so 30 sections of prior context is comfortable.

Rule 2: Force specificity in every section. Generic section prompts produce generic prose. For every section prompt, I include:

• The thesis from the outline
• The 3–5 specific facts/quotes from the NotebookLM briefing to use
• 2–3 specific examples (with names) to draw from
• The tone (see Rule 3)

Rule 3: Hand Claude a voice reference. I paste 800 words of a recent client deliverable or my own published writing and tell it to match the voice. "Match this voice exactly" is the difference between a 6/10 and an 8.5/10 draft. Without it, Claude defaults to a slightly stiff thought-leadership register that readers sniff out immediately.

The actual drafting loop is: prompt → draft → my edits (usually 20–30% of the word count) → next section. By section 3, the voice is dialed in and the edits get faster.

Stage 5 — Fact-Check with NotebookLM (1–2 hours)

This is the stage everyone wants to skip, and it's the stage that determines whether the whitepaper is defensible.

Take every numerical claim, named case study, and "according to" attribution in the draft. For each one, ask NotebookLM:

Look up [SPECIFIC CLAIM] in the source library. Quote the exact text and cite the source. If the claim is not in the sources, say so.

NotebookLM will either:

• Quote the source and confirm the claim (great)
• Quote the source and show the claim is slightly off — usually a number that's been misremembered or extrapolated (fix it)
• Confirm the claim is not in the sources (delete it or replace with one that is)

In the martech whitepaper, this pass caught 11 unsupported claims out of 84 citations. Most were small (the wrong percentage, a misattributed quote), but two were load-bearing — a stat that anchored the entire opening section turned out to be from a 2022 source that had been updated. NotebookLM caught it in 30 seconds. I would not have caught it manually.

This is the part of the workflow that justifies the combo on its own. A whitepaper that can't survive a fact-check is worse than no whitepaper — it kills trust with exactly the people you're trying to convert.

Stage 6 — Polish, Design Brief, and Lead-Gen Packaging (2–3 hours)

The AI's job is mostly done. Mine is:

• Read the full draft in print, mark every weak sentence, rewrite those.
• Cut anything that doesn't earn its length. The first draft is always 15–20% too long.
• Write a 1-page design brief for the designer (or for Canva/Gamma if it's an in-house job) — page structure, callout boxes, chart specs, brand colors.
• Write the landing page copy, the email follow-up sequence (3 emails), and the LinkedIn post promoting the whitepaper. Claude can draft all three from the final document — these are short, low-stakes tasks where the model is at its most reliable.

The promo assets take an hour total. They're the difference between a whitepaper that gets 200 downloads and one that gets 2,000, so I never skip them.

Prompts Worth Stealing

A few prompts I now keep in a personal swipe file:

For finding the angle:

I'm going to send you 5 high-ranking articles on [TOPIC]. Tell me what they all miss, what they get wrong, and what a stronger angle would look like. Then give me 3 candidate angles I could take for a whitepaper that beats the existing content.

For section drafting:

Draft section [N] of the whitepaper, approximately 900 words. Voice: [paste reference]. Thesis: [from outline]. Must include these specific points: [list]. Must use these facts from the briefing: [list with citations]. End with a transition sentence that sets up section [N+1].

For tightening prose:

Read the section below. Identify every sentence that's generic, hedging, or could appear in any article on this topic. Rewrite those sentences to be specific to our argument. Preserve sentences that are already sharp.

5 Traps That Will Eat Your Time

Trap 1: Letting Claude use its general knowledge for "facts." It will produce a beautifully written paragraph containing a stat that sounds real and isn't. The rule: Claude is allowed to know things; it is not allowed to claim things. Every number in the draft must be traceable to a source you uploaded.

Trap 2: Asking NotebookLM to write the whitepaper. It will either refuse, produce a 500-word summary that doesn't qualify as a whitepaper, or — in the worst case — generate something that looks like a whitepaper but is just stitched-together quotes from your sources with no original argument. The argument is your job.

Trap 3: Drafting the whole document in one Claude call. Even with Opus 4.6's 1M-token context, the prose quality drops in the back half. The voice drifts, the arguments repeat, the examples feel recycled. Section-by-section drafting with recap takes more time but produces a document that reads as one coherent piece.

Trap 4: Skipping the fact-check because "Claude cited it in the draft." Claude does not have a "citations" feature for general-knowledge claims. It cannot tell you where a number came from. Only NotebookLM can. If a fact isn't in the source library, it isn't a citable fact.

Trap 5: Not writing the promo copy in the same session. The whitepaper draft is hot in your context window. The landing page, the email sequence, the LinkedIn post — if you don't write them in the same session, you'll either skip them (most common) or write them a week later in worse voice. Bundle them into the workflow.

The Takeaway

A lead-gen whitepaper used to be a 3-week project that required a strategist, a writer, a researcher, and a designer. The Claude + NotebookLM combo turns it into a 2–3 day project for one marketer who can hold the strategy in their head. The economics are absurd — for a 30-page B2B whitepaper, my actual AI cost (Opus 4.6 API + NotebookLM free tier) was under $40, and the time savings compared to a manual workflow were about 80%.

The moat isn't the AI. It's the judgment: which topic to pick, which sources to trust, which arguments to keep, which claims to kill in fact-check, which promo angles will actually pull downloads. The model produces the document. You produce the thinking that determines whether it's worth reading.

If you've been treating whitepapers as a "we'll do it when we have time" project, this workflow is the reason to do them more often. One good one a quarter, with proper distribution, will outperform a year of generic blog posts.