Martin Kowalski (invented)

A long note on how the book was made: voice memos, a journalist agent, a librarian, ideas on a chapter map, and each chapter written by an agent from named sources. It says some chapters are honest placeholders and that notes to Dinis stay in.

Better than most: it admits the AI involvement up front, which answers half my question. But the other half, whether there is a person's thinking in here, is now a claim about a process, and I will test it. Also a lot of newsroom machinery I did not ask about.

The rule is simple: every number, every quotation and every story has to be in a source the chapter names. ✓ verbatim

Question: If the agents write the chapters from the author's articles, how much of the book is his thinking and how much is a rearrangement of it?

Confusion: It says this is not an official version and release candidate 2 came a few hours after candidate 1. I cannot tell how stable any of this is.

Next: Read on, but I am checking the sourcing rule against the chapters that make big claims.

Five chapters before the argument: who it is for, why I am lucky, graphs, where I land, connecting the dots backwards.

A reading order promise. Fine, but five chapters of preamble in a 75,000 word book is a lot.

Question: Why does a book for a founder need a graph chapter before anything about a company?

Next: Go to chapter 1.

Two readers, Gen Z and Gen X, and the name ZX from the Spectrum. The stated objective is that by the end I have started a company, measured by how many companies come from the book. Cites a study of 2.7 million founders. Includes two notes addressed to Dinis in the body.

The ZX Spectrum paragraph landed, I typed those listings in. The stated objective is bold and testable, which I respect. But a book that promises I will have a company and covers the company in a chapter list I can already see is thin, I will check that. The age data suits me suspiciously well.

Our job is to empower the reader to create a startup by the end of the book. ✓ verbatim

Question: What is the 2.7 million founder study, who ran it, and does it say what he says about age 45 and industry tenure?

Confusion: I could not tell whether this is a second edition or a new book; the notes to Dinis say even he has not decided.

Next: Read the next chapter, then jump to the ones that make the claims.

Wonder at what the author can now do, history of earlier technology waves, a dot-com crash paragraph, a pitch that technology finally works for you, and advice to read on paper with a pen.

I am looking for the thinking here and it is mostly mood. The fear essay he cites he could not read past the paywall, and he says so, which is honest but also means he is arguing with a headline. The dot-com line is true and not new. It is an upbeat chapter about feeling lucky, which an unemployed director is not in the mood for.

I could not get past the paywall, so I cannot tell you what it argued. ✓ verbatim

Question: What exactly is the fear he is answering, and who holds it, if the only evidence is an essay he did not read?

Confusion: The line that Gen Z had technology made to hook them and Gen X put the building down is asserted, not shown.

Next: Skim the graphs chapter, then go to the bold chapters. This is not where the book lives or dies.

A graph is nodes and edges, a detective's board, Wardley maps as graphs with positions, nested graphs, and code as a layered graph with a story about an sgit commit reaching nine of 72 commands.

The code-as-layers section is the first idea I have not read in this form, and it is concrete. But it is also about his own tool, and the worked example is his own, unrun, hand-written stories. Interesting, not yet evidence.

That is a way of seeing, not a claim in physics. ✓ verbatim

Question: Does reading a diff at several altitudes find a bug a normal reviewer misses? One commit is an illustration.

Next: Jump to the founder chapter, the biggest claim in the contents.

Expert founders beat the young outsider myth; Horsley Bridge venture return data; software people pay for and dislike; the Nokia inertia argument; explorer, villager and town planner code; and a closing pitch for a code review company built on layered graphs and the author's sgit tools.

The best chapter so far: sourced figures, a stated counter-argument, the Nokia point is a real thought, and he flags where his own articles contradict each other. But the ending is a business plan for his own product line, and the expert founders are two people he helps and himself. The solo-founder statistic and Stripe numbers are the fastest, not typical, which he admits.

I state this so it can be wrong. ✓ verbatim

Question: If the book is evidence the expert founder is the better bet, why are the three examples all people the author is connected to?

Confusion: Whether I should read the closing company as advice or as marketing; it is both.

Next: Now the chapters that are about his products, where I expect the selling.

Zero-knowledge encrypted versioned vaults: sgit, read keys, write-only lanes, and why constraints drive good design. Includes two corrections to his own earlier claims, one about GDPR. Cites the EDPB and the ICO.

This is the strongest technical chapter and I was ready to hate it. The GDPR correction is what a careful person does and I believe it more for it. But it is the author's product, described at length, and the book keeps saying the chapter is not an advert. He does disclose the stake, and says nobody buys it for the encryption.

I have a stake here, sgit is mine, and the market will decide. ✓ verbatim

Question: What does sgit do that an encrypted git remote with age or a cloud KMS does not, in cost and effort to adopt?

Confusion: I lost the thread in the key management section between the vault key, read key, write key, append token and key pair.

Next: Next RiskMandate, the other product.

An agent has reach, a mandate and a gap; barriers are nothing, a prompt rule, a setting the agent controls, or something enforced above the agent; a proposal that vendors accept a behaviour policy as a third input. Aviation and payments liability analogies, then more about running agents.

The reach, mandate and gap framing is useful and I would use it. The claims I know to check: the prompt injection point is right and well cited. The proposal for vendors to accept a policy is a position, not a finding. And it reads as the sales page for a company: prices appear in ch 19, this chapter is the product.

I have a stake in it, and I say so. ✓ verbatim

Question: Who has actually run this on a real deployment other than the author, and what did the gap look like?

Confusion: Whether the five of seventeen rules example is real or invented; it says illustrative.

Next: Check how he handles evidence in the chapter about facts.

Provenance, Popper, 'not found is not false', counting rather than remembering, and how his own memory was corrected by records: nine months was eight, 30 paying customers was ten.

The admission that an investor one-pager said 30 customers when the record says ten is the most trust-building thing in the book so far. Also the part where it explains the book's own method. Some of it is the generation-Z-are-detectives claim, which he labels as his experience.

In the last eight articles when I counted, five of my claims were confirmed, five corrected and two could not be found ✓ verbatim

Question: Is the corrected claim list visible somewhere, so I can see which were corrected?

Next: Go to funding, since that is the thing I am deciding on.

Do not raise before profit; sell existing shares instead of issuing new ones, and lend part back; survivorship bias; an investor selection hypothesis about controllable founders; a UK worked example with tax; what breaks.

Genuinely a thought I had not read in this form, and he hands me the counter-evidence himself: Wasserman, Ewens and Marx, selection effects. The weak spot is the worked example, which is tax and securities structuring presented by a non-lawyer for a company worth 10 million. He disclaims it, but the chapter's whole thrust leans on an untested mechanism.

No study tests this model; the claim is from discipline and incentives, not data. ✓ verbatim

Question: Has anyone actually done a founder secondary plus subordinated loan on these terms, and what happened?

Confusion: How the fork, a possibility, is supposed to be an investor's insurance; the argument is a stretch.

Next: Last stop, the end of the book on how it was made.

Wardley map, the four stages applied to many things, and explorer, villager and town planner as personalities, with a centre of gravity test.

The three-types idea is useful and sells itself. I can say which I am: town planner pretending to be a villager. The Myers-Briggs comparison is his experience, flagged.

Every dot is a claim. The positions are judgements, and the judgements are the argument. ✓ verbatim

Question: Does the book use the badges consistently?

Next: Next.

Connect the dots backwards, multidisciplinary people, a maths-teaching lament, learning with agents, and two of his pivots: Cyber Boardroom and MyFeeds.

A grab bag. The maths-teaching and fish-climbing-tree material is generic self-help and I know the Einstein attribution is false, which he admits. The pivot story is the real content.

Your future self will connect the dots. Your job now is to make dots worth connecting. ✓ verbatim

Question: Is this chapter about careers, learning, or pivots?

Confusion: It tries to be all three.

Next: Skim on.

A one paragraph introduction to the frame.

Fine.

Next: Next.

What generative AI is: buckets of AI, next word prediction, hallucination as a feature or bug, and the Starry Night image that clicked for him.

Plain and mostly right; the next-word explanation is better than most. Printed corrections from the model of his own claims are a nice touch. But I have read this several times.

Most of the trouble people have comes from asking for the second while set up for the first. ✓ verbatim

Next: Next.

A prompt is a lens, a role is a model with a description, an agent is a loop with a goal, and tools changed everything.

Sound, and the caveat that an agent is not just a loop is good. Familiar to anyone who has used coding agents for a year.

Good agents add planning, checks and limits; the loop is the skeleton, not the whole animal. ✓ verbatim

Next: Next.

Everyone who can brief an agent is a developer; spreadsheet error stories; Portuguese as the sixth most powerful programming language.

The spreadsheet horror stories are sourced and correct. The 'citizen developer is good news' claim skips that IT does have to clean it up. Repeats 'Execution just got cheap' from ch 1.

The citizen developer is the spreadsheet grown up. ✓ verbatim

Question: Who maintains what citizen developers build?

Next: Next.

Explorer coding versus vibe coding, why explorer code is not ready to be in line, and the handover.

The core distinction of the book and it is good. It also appears in ch 4, 7, 17 and 18; by now I have read it four times.

Explorer coding is not a way to ship. It is a way to find out. ✓ verbatim

Question: What is the handover in practice: a checklist, a contract?

Confusion: The handover is promised but remains a question to Dinis at the end.

Next: Next.

Short introduction to the job-hunting part.

Aimed at a different reader from me.

Next: Next.

Hired for the workflow: running agentic workflows is the competitive advantage, with a history of past rare skills.

The pattern of rare skills becoming normal is true. The assertion that the one answer in 2026 is running agentic workflows is a prediction presented as a conclusion.

Each of those skills later became normal. ✓ verbatim

Question: What would a hiring manager actually test for?

Next: Skim.

Act as the job you want: do the work in public before you are hired; an unfinished section on the 2026 job hunt.

Reasonable advice that predates AI. The last section is an admitted placeholder, so the chapter ends on a hole.

The work gives you the CV you did not have. ✓ verbatim

Question: What about the interview process now?

Confusion: The placeholder section stops with nothing.

Next: Skim.

Publish constantly; his Leanpub books stall; the voice memo to checked article loop; 148 messages and 65,323 words over 27 days.

The admission that his books stall is good. The recipe is concrete. But it is a chapter about running a newsroom, which is not obviously what a founder needs.

Small finished pieces beat big batches. ✓ verbatim

Next: Next.

A short introduction to working with agents.

Fine.

Next: Next.

A Google Drive and Claude workspace: one folder per contact with five documents and six lines of agent rules.

Practical and I like the Ledger, add-only. But it is a tutorial in the middle of the argument, and I see in the contents that it reappears as ch 30.

In a spreadsheet the record is the row, and the row is a summary you typed. ✓ verbatim

Question: Why is the same workflow in two places?

Confusion: I could not tell the two chapters apart.

Next: Next.

Agents as sessions with focus, one session per topic, writing lessons and debriefs, the brief as the unit of work, and a daily brief from a librarian agent.

Good practical content and matches my own experience. A few lines of this are the same as ch 7.

A thread is the model's frame of reference, and everything in it is context for the next answer. ✓ verbatim

Next: Next.

A short introduction to direction and judgement.

Fine.

Next: Next.

Direction is the scarce skill: places of highest leverage, and knowing when to stop.

The GUID randomness meeting is a true thing I have sat in. The core claim is reasonable and old. Someone who has managed engineers knows it.

Agents can build in any direction; what they cannot generate is the sense of what to build next, for whom, and when to stop. ✓ verbatim

Next: Next.

The centre of the edition, with a legal and tax disclaimer built in.

The disclaimer is sensible.

Next: Next.

Ship it, give it away briefly, take it away and ask if they miss it. Flags a tension with his own vibe coding wording.

Simple and a decent test, though not new: it is a form of the Sean Ellis test. The tension flagged with his own earlier article is honest.

If they shrug and go back to their day, you do not have a product yet. ✓ verbatim

Question: How many people is enough for the signal?

Confusion: The product testing story is told in two places.

Next: Next.

Four moves: define, price, deliver at linear cost, offer free to people who know you. RiskMandate's price ladder as the example.

It uses RiskMandate throughout as the example. It is the author's product and the early access programme. The review date in December is in the book, which dates it.

Price it first, because the price is the claim. ✓ verbatim

Question: Is the example the product's real performance or an illustration?

Confusion: The chapter closes with a review that has not happened yet.

Next: Next.

Sell by answering a question already in the room, and value that lands the day they switch, with a day on digital sovereignty.

The point that sovereignty and cost are important but not urgent is a real insight I have seen on registers. Again RiskMandate is the running example.

The reason to move must never be sovereignty, and it must never be cost reduction. ✓ verbatim

Question: What is the evidence beyond one conversation with one operator?

Next: Next.

A placeholder: the company in the UK, Portugal, Lithuania and the EU is not written, with a section on hiring through Upwork.

This is the chapter the book's objective depends on, and it is a placeholder with two lines. In the sold book this must be written or the promise has to go.

This chapter is a placeholder, and it says so. ✓ verbatim

Question: Where is the actual how-to?

Confusion: I expected to be told how to form a company and got a note.

Next: Skip on.

A short introduction to the serverless part.

Fine.

Next: Next.

What serverless means: an idle-cost table dated 18 October 2026 and the 'ephemeral town planner services' name.

The idle-cost test is sharp and useful: if it bills by the hour idle, it is not serverless. The table is a snapshot that will age. Mild whiff of renaming an established term.

I would rather replace the word than keep defending it. ✓ verbatim

Question: Who verified the prices?

Next: Next.

Open source: what it is not, the licence misconceptions, and the fork next door as investor insurance.

Mostly fine. The 'moat is a rate, not a wall' argument is an assertion. The 220-fork study with 7 merges is a real number.

Eight years on, everything in that chapter still holds, and the reason to do it has become clearer to me ✓ verbatim

Question: Does the fork next door work for a company whose customers are enterprises?

Next: Next.

Shipped means the user has it: a release that pushed cleanly and never reached users for forty minutes, and the fix of asking the live site what version it serves.

A concrete incident with a concrete fix, I like it, though it is a dev-ops chapter for people who already ship. The five-users plan at the end mentions a study; it is thin.

The checks were not weak. They were checking the wrong boundary. ✓ verbatim

Next: Next.

A short introduction to the long game.

Fine.

Next: Next.

Backup your life, graphs as memory, edges must be verbs, supersede never delete, and keeping a graph rather than getting it right.

The rules about edges are good practice. Repeats ch 3 and 16 on graphs and provenance.

You do not have to get the graph right. You have to maintain it. ✓ verbatim

Confusion: I found the graph material repeated across ch 3, 16 and 28.

Next: Next.

Civic responsibility: his 2010 and 2016 Lisbon talks, Hacking Portugal, what came true, and the state's plumbing.

Honest about what did not happen, which is rare. But this is a Portugal-specific chapter in a book for founders anywhere.

What I got wrong was the mechanism. ✓ verbatim

Question: Who is the reader outside Portugal?

Next: Next.

Eleven hands-on workflows dated 2026.

A good idea to separate the dated material.

Next: Next.

The same folder per person and six lines workflow as ch 13, in short form.

Close to a duplicate of chapter 13.

The difference is not the tool. It is what counts as the record. ✓ verbatim

Question: Why is it here and in ch 13?

Confusion: Duplicate.

Next: Skim on.

Build a first graph by hand in seven steps, starting from a question.

A real exercise. The warning not to pick a subject instead of a question is good practice.

A graph has no front page. ✓ verbatim

Next: Next.

Explain one idea five ways with a model, with the same lens argument as ch 7.

An hour exercise that repeats the argument it is practising. Thin.

It stops being magic. That is the point of understanding it. ✓ verbatim

Confusion: Repeats ch 7 including the maths example.

Next: Next.

Run an agent loop by hand with yourself as the harness.

A nice exercise for a newcomer; I know all of it. The final memory line is good.

Nothing is remembered by the model between sessions. Everything is remembered by the files. ✓ verbatim

Next: Next.

Storyboards as data drawn by Rough.js, a cast file and one file per story.

This is the site's own tooling and not what a founder needs. It reads as the author's project.

There is no image file to edit. Change the data and the picture changes. ✓ verbatim

Question: Who is this for?

Confusion: I could not tell why a founder would build this.

Next: Next.

Meeting briefs and follow-ups; says it is short on purpose because the sources are not there.

Another admitted placeholder, and it repeats ch 16.

This is a short chapter, on purpose. ✓ verbatim

Next: Next.

The recipe for a voice memo to checked site and an agent-worked inbox that never sends.

Concrete steps, including the rule that a failed check stops the release. Much of it appears in ch 12.

The house rules are the product. ✓ verbatim

Confusion: Overlaps ch 12 and ch 13.

Next: Next.

Form your company: not written, with what is known and what to decide first.

Second placeholder for the same promise as ch 22.

So this is the honest version: what the chapter will cover, the little the sources already say, and the decisions you can make before the steps arrive. ✓ verbatim

Question: Why two chapters for a missing one?

Next: Next.

A table of cloud and AI startup credit programmes as of 7 October 2026 with amounts and eligibility.

Useful and the most checkable page in the book, and it will be out of date in months. It says so.

A year from now several rows of the table above will be wrong ✓ verbatim

Question: Who pays for the book to be renewed each year?

Next: Next.

Price first, offer to people who know you, stop rule before the first message, count the results.

The same four moves as ch 19 in the same words, with the same RiskMandate example.

The research on paying first is not kind ✓ verbatim

Confusion: Repeats ch 19.

Next: Next.

A weekend build: static host, data in a vault, a model on the page, a payment link.

The most specific chapter and admits the 83 lines of server are an exception and have not run on a real account. Good honesty, weak foundation for a promise of a product by Sunday.

It has not run on a real account, does not do files over 4 MB and has one shared access token. ✓ verbatim

Question: Does a stranger actually pay through it?

Confusion: The product uses the author's vault service.

Next: Next.

Introduction to the process part.

Fine.

Next: End.