Book, course, or writing agent: a SaaS founder triage

It's a Sunday in Den Haag. A founder we work with has three tabs open: a Notion doc with 47 essay topics, a half-edited course intro recorded in March, and the HN thread asking whether AI has already killed self-help nonfiction. Their SaaS does €4.3M ARR, mostly through one newsletter at 8,400 subscribers. They want to know which of three things to ship next: a second book, a 40-video YouTube course, or a finetuned writing agent that handles the long-form backlog. They want a method, not a vibe.

This is the scoring method we use with them, and with the four other Dutch founders running similar shops who pinged us the morning after that thread hit the front page.

The three options are not equivalent bets

Founders frame this as a choice between formats. It isn't. It's a choice between three different bets on where the next twelve months of attention live.

The book is a credentialing bet. It moves you from "person who has a newsletter" to "person who wrote the book on this." The course is a distribution bet — YouTube's recommendation engine will either feed you for two years or ignore you in week three. The writing agent is an inventory bet: you trade a one-time training cost for a steady supply of essays, landing pages, and replies you can ship without opening a blank document.

You can run two of these in parallel. You almost never want all three. The founder who tries all three ships none.

The three things we score on

We don't score on what feels exciting at the kitchen table. We score on three things we can measure or argue from evidence: reader retention at the existing 8,400 newsletter subscribers, AVG-defensible source attribution, and who edits the artefact when the niche shifts in a quarter.

Each option gets 0, 1, or 2 on each axis. Highest total wins. Ties go to the option the founder will actually finish — which is the option whose first week of work doesn't make them want to lie down.

Reader retention at 8,400 monthly subscribers

This number matters more than which format is trending. A book at a list of 8,400 is a slow-release object. If 6% of the list pre-orders and another 8% buys in the long tail, that's about 1,200 readers in year one. Enough to seed reviews, not enough to repay the year it took to write. The book scores on retention only if the existing list reads to the end — and across the founder newsletters we audited in 2025, the welcome-email read-rate by week four sat between 22% and 31%. That's the ceiling.

The 40-video course pulls a different number. YouTube watch-time on a niche founder channel under 10k subscribers is brutal for the first 15 videos. We've watched dev-tools channels sit at a 40-second average view duration for three months before the third or fourth video lands. The course scores well on retention only if the founder already has 1,000+ subscribers or is willing to spend €18-25k on a video editor for twelve months.

The agent scores on retention differently — it doesn't pull new readers, it keeps the existing ones reading. Three essays a week instead of one a month is the difference between a list that opens you 38% of the time and a list that opens you 11%. We've watched the same founder's open-rate halve in two months when they switched from weekly to monthly. The list doesn't unsubscribe, it just forgets.

Scores on retention

• Book: 0 if the list is under 12k and read-rate is under 30%. 1 if either passes. 2 if both.
• Course: 0 unless YouTube subs are above 1k or the founder has €18k+ for an editor. 1 if either. 2 if both.
• Agent: 1 by default. 2 if the founder currently sends less than weekly.

AVG-defensible source attribution

This is where most founders flinch. If you publish under your name, your sources need to be yours or correctly attributed. Under the AVG, the EU AI Act's transparency clauses, and the broader copyright case-law that's been building since 2024, you are on the hook for what your name signs. "The agent wrote it" is not a defence the Autoriteit Persoonsgegevens has ever accepted in writing.

A book lets you cite. You have endnotes, a bibliography, an editor who checks them. Score: 2.

A 40-video course is harder. Most founders we've watched ship a course never put sources on screen. They reference research verbally and viewers don't pause to verify. If a claim is wrong, the LinkedIn post about it spreads faster than the correction. Score: 1, and only if the founder commits to on-screen citations in every video.

A writing agent depends entirely on whether it was built to log sources. A bare LLM that paraphrases its training set is a legal liability — you cannot point to where a claim came from, which means you cannot defend it. The pattern we ship for clients in regulated niches is simple: every agent draft has to cite rows in a source table, and the publish button is disabled if any sentence in the draft cannot point at a row.

-- every agent draft must cite rows in this table
create table sources (
  id uuid primary key default gen_random_uuid(),
  kind text not null check (kind in ('case','statute','article','internal_doc')),
  url text,
  cited_text text not null,
  added_by text not null,
  added_at timestamptz not null default now()
);

create table draft_citations (
  draft_id uuid not null,
  source_id uuid not null references sources(id),
  sentence_hash text not null,
  primary key (draft_id, source_id, sentence_hash)
);

Score: 2 if you build it that way. 0 if you don't.

Who edits the agent when the niche shifts

Niches move. Keyword distribution in a SaaS vertical changes every quarter, sometimes faster after a model release shuffles search behaviour. The HN thread from last week is the obvious example: "self-help nonfiction" became a contested term overnight, and any agent trained on last quarter's distribution writes around the wrong corners.

The book is locked the day it ships. You can write a second edition, but most founders don't. Shift-resilience score: 0.

The course is partially locked. You can re-record a module in a weekend, but you'll never re-shoot the intro. After two distribution shifts (six months, in fast verticals) the course feels dated and watch-time collapses. Score: 1.

The agent is editable, in principle. The question is by whom. If the founder is the only person who knows what's true in the niche, and the founder can't touch the agent's prompt or sources, the agent becomes a liability the moment the niche moves. We've taken over three projects in 2025 where the previous team had built a beautiful RAG system that no one on the client's side could update without a JIRA ticket and a six-day turnaround. By the time the next draft shipped, the niche had moved again.

The agent scores 2 only if the founder (or a single non-engineer on the team) can do all three of the following without opening a deploy pipeline:

• Add a new source to the library by uploading a PDF or pasting a URL.
• Rewrite the system prompt and see the next draft change within five minutes.
• Read the agent's source attributions and approve or reject before publish.

If any of those needs an engineer, drop the score to 1. If two of them do, drop to 0. We've never seen an agent score 2 here unless it was designed for non-engineer editing from day one.

How the Den Haag founder scored

Newsletter at 8,400 with a 34% read-rate. No YouTube channel. No engineer on call. One existing book, three years old.

                  Retention   Sources   Shift-edit   Total
Second book           2          2          0          4
40-video course       0          1          1          2
Writing agent         2          2*         1*         5
* contingent on building it the right way

The agent wins on points, but only if it is built so a non-engineer can swap sources and rewrite the prompt without a deploy. The book is the safe second pick. The course doesn't score — not because courses are bad, but because this founder lacks the YouTube position and the editor budget.

When we still pick the book

The method is not a religion. There are three cases where we tell the founder to write the book even when the agent scores higher.

One: the niche moves slowly. Tax law, accounting workflows, classical-language tooling — these move on a three-to-five year cycle and a book stays accurate. Two: the founder sells to enterprise procurement teams who will not buy from someone without a hardcover on their shelf. Three: the founder has already written one book and the second one is the real leverage. The first book was credentialing; the second is positioning.

None of these were true for the Den Haag case. So the agent it is.

The smallest thing to do today

Open a fresh doc. Write down your newsletter open-rate, your YouTube subscriber count, and whether anyone non-technical on your team can edit a system prompt without filing a ticket. That is the entire input to the method. If those three numbers don't justify the format you've already started building, stop building it before the next sprint.

When we built the writing agent for a Rotterdam legal-tech client, the thing we kept hitting was the source-attribution problem — every essay needed to point at a row in their case database, not at a fuzzy memory of one. We ended up treating the agent as a query layer over their existing knowledge graph, not as a generator that occasionally cited. If you're somewhere in this triage and the AI agents column is where you keep landing, that's the part to figure out first.