← Back
Letter 0 · 1 May 2026 · 3 min

Using NotebookLM to give the cast a voice

How I'm building fifteen characters from a research substrate, and why this is itself a demonstration of one of the book's arguments.

Explore this page

You’re getting this because you signed up to follow the build of The Weight of Small Decisions. Until now, the only thing that has existed is the page at pramod.ch/book. This is the first letter that adds anything new on top of it.

I’m midway through Step 6 in the public roadmap on pramod.ch/book: the fat outline. Twenty-four chapters, each one specified before any of them gets drafted. Concept, characters present, location, emotional beat, the unresolved question that closes the chapter. It feels harder than the drafting will be.

The part of the build worth sharing first is the one that has been most useful and is hardest to talk about without sounding like AI hype. So let me be specific.

Using NotebookLM to give the cast a voice

The cast has fifteen named characters. Six of them carry weight across the book. Each of those six has a profile, but a profile is not a voice. A voice is what comes out when the character is asked something they did not expect.

What I have been doing: for each of the six, I built a small NotebookLM notebook seeded with the substrate that would have shaped that person.

For the protagonist, an operations executive at a mid-size pharma, that is regulatory filings from cell-and-gene-therapy programmes, transcripts of S&OP review meetings, oral histories from operations leaders who came up through CMC, hiring criteria for her role, and the books she would have read at three career inflection points.

For the CFO, a different stack: capex memos that actually got signed at biotech boards, finance-leadership podcasts, board-pack templates, the Helmer 7-Powers tradition.

For a contract-manufacturing partner running a sterile fill line, a third: technical agreements, deviation reports, oral histories from CMOs that have lived through real failures.

Then I interview them.

“A Form 483 just landed at your sister site in Singapore, with three observations that touch processes you also run. The next board update is in nine days. Walk me through what you actually do today, hour by hour.”

The answers are not what I would have written if I had been making it up. They reflect the substrate. The CFO’s first instinct is not the protagonist’s. The COO who came up through CMC frames the situation differently than the one who came up through commercial. Their disagreements have texture.

I am not asking the model to write the book. The decisions, the prose, the scenes are mine. NotebookLM is doing the thing I cannot do at scale: simulating what fifty practitioners in each role would say if I could interview all of them. The output is a corpus of authentic-feeling reactions that I read, argue with, sometimes throw out, and occasionally lift a single sentence from.

This is, I notice, a methodology demonstration of one of the book’s own arguments. AI is the diagnostic lens. It is not the hero of the story. It surfaces what would otherwise stay tacit. A human still has to decide what to do with it, which is exactly the novel’s claim about decision architecture: better systems do not remove judgement, they make judgement legible and accountable under pressure.

What’s coming

The fat outline lands in a few weeks. After that, two sample chapters drafted as voice tests: one narrative chapter and one concept interlude (a short essay sitting between chapters). Then the proposal.

The next letter from me is the one that says “the fat outline is in,” and it will share what the structure ended up looking like.

If you have specific questions about how the build is working, or scenarios you would want to see characters tested against, hit reply. Replies come straight to me.

Pramod