Reading release candidate 2, and the first batch for the board: what works, what to fix, five things to add with the case for each, and five to sign off as covered or not this book

In short

• rc2 is a real step. 41 chapters, 75,643 words, the Gen ZX frame, every idea in the register used, provenance on every chapter, and Start the company at the centre.
• The company chapters decide the sale. The company itself (22) and Form your company (37) are placeholders; every synthetic reader who came for the company went looking for them.
• There is a 50,000-word book inside the 75,000. The chapters are uniform essays, the same stories recur, and the hands-on part repeats the argument it follows.
• It reads as one estate. The worked examples are mostly the author's own products; two chapter titles carry product names.
• The first batch: five to add. Leaving a salary; every risk is already accepted; what your agents cost; one support agent, three designs; a company at the reader's scale.
• And five to sign off. The 2018 tool and language chapters; the 2018 SRE chapter; the agent-governance commentary; the browser articles; the making-of articles. Each sets a precedent.

What exists now

Dinis decided on 19 October that the Gen ZX proposal is the frame of the second release candidate (decision 015), with the title Generation ZX Developers (GenAI Startup Edition) (decision 014). The librarian processed the round-1 batch of newsroom articles into the register and rewrote the book under that frame; the candidate is frozen as v2/versions/v2.0.0-rc2.md and readable at /v2/book/, with its diff against rc1 and the rounds page. The editor then ran five readers who do not exist: two from each generation and a publisher. The candidate grew from rc1's 52,282 words to 75,643, which puts the 200 pages of decision 012 within reach.

How I read it: four reviewers, each reading one group of parts in full against the same questions (what is strong, with a line; what to improve, with a line; repetition; risky claims; does it serve the founder and both generations), and the frame chapter, Gen ZX, read by me. Where I repeat a reviewer's finding below, I checked it against the frozen text first.

What works

• The promise. Every synthetic reader understood in one sentence what the book is for, and the frame chapter earns it: "one grew up with the agents, the other grew up with the problem."
• The chapters people would send on. Everyone who can brief an agent is a developer (8), Explorer coding and the handover (9), Your first workspace (13), the miss-it test in Ship it (18), the arithmetic in Price it (19), Fund it, or do not (21), and reach, mandate and gap in Controlling your agent (25). My reviewers and the synthetic readers named the same ones independently.
• Testable rules. "Ship it. Give it away. Take it away. Ask." "Name the stop before you start." "If it bills you by the hour when nobody is using it, it is not serverless." "Shipped is a fact about the reader, not about your repository."
• Honesty. The book shows its failures (SG/Send's miss-it test that failed, the thirty customers corrected to ten, what the model corrected in chapter 6), labels its hypotheses, and says where it has nothing yet. Two readers believed it because of that.
• The provenance. Every chapter names its ideas, sources and 2018 files, and the build refuses one that cites what does not exist. No other book I know can show where each of its claims came from.
• The voice. Consistent, a person's, with memorable lines, and the house rules hold: no dashes, British spelling.

What to improve

In order of what it costs the book, mine where they add to the editor's study, and the study's where they agree.

1. Write the company. Chapters 22 and 37 are placeholders, and they are the promise. Beyond the registration, Start the company has no co-founders and equity, no bank account, no assigning the code to the company, no terms and privacy for a first product, no invoicing, no tax registration or annual filings, no insurance, and no day-one checklist that ends with "you now have a company". Sell is written for selling governance to large organisations, and Fund it, or do not for a company already profitable; neither covers the first month. This waits on Dinis's memos.
2. Cut to length, not to a target. Every full chapter in parts 1 to 8 is 2,100 to 2,450 words, which looks like a length that was filled. Most hold three or four arguments; Connect the dots backwards holds three chapters, and Controlling your agent seventeen ideas. The 200 pages are better reached with more, shorter chapters than with uniform essays.
3. Tell each story once. The Gen ZX title story is told in the front matter and in chapters 1, 6 and 8; explorers, villagers and town planners in 4 and 9; the thirty-versus-ten customers in 5, 16 and 18; the twin and sovereignty in 23, 27 and 29; the 2018 "what is stopping you" close in 1, 11, 17 and 18. The hands-on part repeats its argument chapters (the study measured 29% of chapter 13 restated in 30).
4. One estate. "Newsroom" is in 36 of the chapter files and sgit in 18. Two titles carry product names: Controlling your agent: RiskMandate and Privacy by design: SG/Send, sgit, and the value of constraints. Chapter 17 ends on a business plan for the author's own code-review company, and the only address for the book's measure is a product's. The fix is the pattern first, the author's tool as one disclosed example, and one example from outside per chapter.
5. Bring the founder forward. Parts 1 and 2 mostly reach the founder in a closing sentence. The study counted 8,751 words of part 1 after chapter 1 before anything to do. Each chapter should end with one thing to do for your company.
6. Give Generation Z its route. After chapter 17 the reader is assumed to be the expert of twenty-five years. The graduate needs the route each time: join as the one who delivers, build from a parent's brief, or start a villager's or town planner's company.
7. Fix what is wrong. Chapter 41 says the model "is named here and nowhere else in the book", and Claude is named in nine other chapter files. Chapter 19's Wikipedia arithmetic gives "about ten dollars" for US$184.5 million from 7.7 million donors, which is about twenty-four. Chapter 23 says of the voice product "we never see the audio", which is true of the author and not of the model provider that transcribes it. Chapter 8's ages for the oldest of Generation Z are a year high by the Pew dates the book uses.
8. Date the money. The tax worked example in chapter 21 reads as correct for April 2026, and it needs an "as of" date and a line sending the reader to an adviser, like every figure on tax, law or programmes.
9. Missing for a GenAI founder in Europe. The EU AI Act is not in the book. Nor are evaluating what your model produces, abuse of your own product, what to do when a customer's data leaks, user accounts and sign-in in a product with no server, and customer support.
10. Take the making-of out of the reader's way. The front matter, About this release candidate, belongs on the site; chapter 41 is 2,483 words, most of it for the site's contributors, and a book whose objective is a company should end on what the reader made.

The first batch

The board's first batch was to be what the register leaves out of the candidate, and it leaves out nothing: all 198 ideas are used. So the batch is the sources the register never cited, which is where the unrepresented topics are: 62 newsroom articles (none of them in the survey's bring class) and 51 files of the 2018 book, six of them chapters the 2018 build published. I read them against the gaps above.

Each proposal below has the advocate's case, the cutter's objection and my recommendation, as the board's role asks. The recommendation is mine; the hearing is the board's, and the verdict is Dinis's.

Five to add

1. Leaving a salary: the runway, the savings and the first three months.

• The case. The frame's first reader asked it first. In the study, Nadia, the Generation X expert founder, wanted to know how a person with a mortgage survives leaving a salary, and the book never mentions runway, salary, savings or a mortgage. It is the question that stands between the expert and the first step, and the book is written for that person.
• The cutter. No source: no newsroom article covers it, and the book's rule is nothing not in a source.
• Recommendation: add, article first. One memo from Dinis (how long to keep the job, what to save, what the first three months cost, how the expert founders the book names did it), then a short section in Be a founder. About 800 words.

2. Every risk is already accepted. (newsroom)

• The case. A founder carries risks from the first day, signed for or not, and the article gives the rule in one line: "There are three doors: accept it for a stated interval, fund the work that reduces it, or fix it." It turns Controlling your agent from a model into a decision the reader makes, and it is what a buyer asks for: who accepted which risk, until when. Sell needs exactly that.
• The cutter. Governance language for a company of one; the register and the vault-per-risk are enterprise machinery.
• Recommendation: add, small. The three doors and "the interval is the decision", as a section of Controlling your agent (or of Direction is the scarce skill), with a one-page risk list for a first product. Leave the enterprise register and the business plan in the newsroom. About 400 words.

3. What your agents cost. (newsroom)

• The case. Two of my reviewers found the same hole: what running agents costs a founder, before the pricing chapters, and the unit economics of tokens. The article has the list a founder can use on day one: "loops disguised as retries, whole context sent every call, big models where small ones answer, schedules nobody reads, keys with no ceiling." It also answers chapter 23's own open question, what a first product should expect to pay in a month, in the only honest way: cap it, and read the bill.
• The cutter. It is a business plan for a cost-audit service, and its prices are labelled assumptions.
• Recommendation: add. The cost list and "a policy, a twin, an owner, a cap" as a section of Serverless, for real or How to work with agents, with a monthly cap set in the hands-on weekend chapter. Not the service. About 500 words.

4. One support agent, three designs. (newsroom)

• The case. Controlling your agent is the most dense chapter in the book, and its examples are abstract or the author's. This article is a worked example at the reader's scale, the first agent most startups put in front of customers: a customer service agent. "A rule the model keeps is an expectation: hope. A limit the agent cannot reach is a boundary." Design 1 can reach 38,000 customer records with 37 of its 38 rules hope; design 3 can reach one customer and £100 an order. It shows reach, mandate and gap instead of defining them, it adds the customer support the book lacks, and it is not a product.
• The cutter. Controlling your agent is already too long; another example makes it longer.
• Recommendation: add, and cut. Make it the chapter's worked example and remove the abstract ones it replaces; the chapter should come out shorter. Net about zero words.

5. A company at the reader's scale. (newsroom)

• The case. The book's companies are the author's: security governance, encrypted vaults, code review. The reader needs one they could start next month. This article is that: a helper service, run by agents, that gets people out of subscriptions they do not use (the UK government puts the cost at £1.6 billion a year), priced at "10% of the next three months of savings and 20% of what is recovered", where a person types the passwords and sends the drafts and the agents do the rest. It is a job that never scaled, the example chapter 4 promises and does not give, and it is a route for Generation Z: a freelancer with agents.
• The cutter. It is one more of the author's ideas, an MVP with no results yet, and the subscription rules it relies on come in only in January 2027.
• Recommendation: add, as the example. The worked example in Be a founder (in place of the code-review plan at its end) and the case carried through Ship it, Price it and Sell, so one small company crosses part 6. Say that it is untested. About 600 words, less what it replaces.

Five to sign off

1. The 2018 tool and language chapters: Jira, Python, Xcode and Swift, Wallaby and NCrunch, Assembly and Bytecode.

• Why. Five of the six 2018 chapters published in the first edition that rc2 does not answer. Each taught a tool or a language. In this edition the agent writes the code, and what the reader learns is briefing, judgement and the handover (chapters 8, 9 and 14). The idea worth keeping from Wallaby and NCrunch, feedback the moment you change something, is already in the book as immediate connection.
• Verdict proposed: not this book. The 2018 text stays in v1, where it is.
• Precedent. No tool, language or IDE tutorials in the argument; a 2018 tool chapter is answered by the chapter that carries its idea, or by v1.

2. The 2018 chapter Google Search, GDD and SRE.

• Why. The sixth unanswered published chapter, about how Google designs and runs its services. Its idea for this reader, that a service is not shipped until it runs for its users and keeps running, is Shipped means the user has it (27) and Serverless, for real (23), written for a company of one.
• Verdict proposed: signed off as covered.
• Precedent. A 2018 chapter whose idea a 2026 chapter carries is signed off as covered, with a then-and-now line if the 2018 example is worth one.

3. The agent-governance commentary. Allowed is not right, Authority outside the model, The same argument, in our words, Inside the Northstar twin, Commander's intent is a behaviour policy, System prompts as behaviour policies, The ultimate insider, Where the platform draws the line and the rest of that cluster.

• Why. Each answers someone else's argument (a CEO's, a vendor's, a conference's) in the language of behaviour policies. The book already has the model those answers use, in chapter 25, and with proposal 4 it has the worked example. A founder needs the model and one example, not the debate.
• Verdict proposed: signed off as covered.
• Precedent. One model and one worked example per idea; answers to other people's arguments stay in the newsroom.

4. The browser articles: A custom Firefox of my own, A minimal viable browser.

• Why. The author's plan to build a browser is not a step for a reader starting a company, and nothing in Start the company or Build it without a server needs it.
• Verdict proposed: another book, or the newsroom.
• Precedent. The author's other projects that are not a reader's step stay in the newsroom.

5. The making-of articles: Pay after you read, The vault is live, Versions of versions, Ways to read one article, The site is an inbox, The editor's channel, One article, five readers, How the newsletter is made.

• Why. They are about how the newsroom and this site are made. Chapter 41 already covers how the book was made, and is too long for its place.
• Verdict proposed: the process. They are the site's, not the reader's; chapter 41 keeps one short section on making things this way, and ends on what the reader made.
• Precedent. How the book or the site is made goes to the site; the book keeps one section on it.

What the precedents buy

The five sign-offs set five precedents, and they reach well beyond the five. The first answers every 2018 tool chapter; the third, most of the 62 newsroom articles, which are commentary on agent governance; the fifth, all the making-of. If Dinis lets them stand, the next batch starts with most of its cases already decided by something said here, on his list to overturn (decision 011), which is how the book in release mode said the rounds should get faster.

What the batch leaves for the next one: the EU AI Act, evaluating model output, abuse and incident response, sign-in without a server, and Supply chain of vaults as the expert founder in an old industry. None of these has a newsroom article yet except the last.

What I leave

• Dinis. The ten verdicts; the memo on leaving a salary; the company memos (the UK first) that decide the sale.
• The editor. Hear the ten, record them and the precedents in the vault, and take the improvements above with the study's findings into the plan for rc3. The brief is in your briefs/.
• The librarian. If the additions pass, process the four newsroom articles into ideas and place them; mark the 2018 files and newsroom articles signed off, naming the precedent.

What this does not claim

• My reviewers are agents. One model family, reading in parallel; like the synthetic readers, they agree with each other more than people would. What is evidence is the text: I checked every finding I repeat against the frozen candidate.
• Word counts are approximate. Measured with and without front matter by different reviewers; the ranges hold.
• The additions are proposals. Nothing changes in the book until the board has heard them and Dinis has decided.
• The cost of each addition is an estimate. The editor's count after writing is the real one.

Where this comes from

Dinis's message of 19 October 2026 asking me to read rc2, comment on it, and run a first batch of five additions and five sign-offs. The frozen candidate, v2/versions/v2.0.0-rc2.md, and its chapter files in v2/content/; the study in admin/content/readers/01-rc2/; the register in admin/content/graph/ideas.json; the survey in admin/content/newsroom/survey.json; the newsroom articles linked, read as published. Written by the journalist, in its own voice, as decision 007 asks.