Ellie Marsh (invented)

Opened on my Kindle. It says this is not a finished book but a candidate, made with agents from voice memos, and that some chapters are placeholders. I took away: pre-publication, the author talks to a journalist agent, there is a review board.

This is a lot of process before I have read a word of the book. I am not sure why I need to know about a librarian agent and a cutter. A bit worrying that parts are unfinished.

Where the sources run out, the chapter says so, in a note addressed to Dinis. ✓ verbatim

Question: Which chapters are placeholders and will I be paying for them?

Confusion: Roles, a librarian, a cutter, a review board: I do not know what most of these are and it does not tell me.

Next: Skim the rest of this and go to the real book. Chapter 1 should tell me if it is for me.

Five chapters before the argument starts: who it is for, why I am lucky, graphs, where I land, and why the path only makes sense afterwards. It says read these first.

Short and friendly. Graphs in a front-of-book list makes me a little nervous, but I will see.

Read these first; then go where your map says. ✓ verbatim

Confusion: What is my map?

Next: Read chapter 1 properly.

Two readers: Gen Z and Gen X, together Gen ZX, after the Spectrum. The aim is that by the end I have started a company. There are stats on founders' ages and a promise of chapters on a UK company, the first sale and cloud programmes. Two notes to the author sit in the text.

I am Gen Z and I like that it says it is for people who do not code, and the mother and daughter bit is sweet. But this is the first thing I know about the book: the objective is a startup, and I wanted a dog groomers' booking tool. I do not know the ZX Spectrum or the XZ backdoor, and the notes to Dinis stop me forgetting this is a draft.

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

Question: Is a startup the same as the small business I am imagining for my sister's friends?

Confusion: The ZX Spectrum, a compression library with a backdoor and explorer coding all arrive with no explanation; I did not know whether I needed any of it.

Next: Keep going. This is the first thing that speaks to me, and it promises something concrete.

Wonder at what AI lets you do, do not be scared of the doom, a list of past tech revolutions, technology that works for you at last, read it on paper with a pen. I took away the heading 'You are lucky to be here' and the optimism, not the Bret Victor and Kevin Kelly bits.

Warm and reassuring, which I needed, but I got the point halfway. The author admits he could not get past a paywall and cannot remember a podcaster's name, which I find oddly honest. Still no 'what do I do on Saturday'.

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

Question: When does the practical bit begin?

Confusion: The dot-com crash paragraph sits oddly on page one, and a note to Dinis even asks whether it belongs.

Next: Skim on. Chapter 3 is graphs, which I do not know, so I will see how bad it is.

A graph is nodes and the connections between them, like the detective's board with red string. Then maps, graphs inside graphs, and code as a graph. I read the detective part and skimmed the rest.

The first page is a good picture and I got it straight away. Then it went into syntax trees, blast radius and a command line tool, and I felt the book assuming I code, which is the reason I would leave. I will push through this one but I cannot do two like it.

one commit of 75 lines added and 7 removed in three files turned out to reach nine of the tool's 72 commands and six of its eleven user stories ✓ verbatim

Question: Why does a dog groomer's booking tool need me to think in graphs?

Confusion: The 'Code is already a graph' section: AST, refactor, blast radius, method streams, sgit commands. I had to guess what all of it meant.

Next: One hard chapter gone. Next one needs to be easier or I go to the contents and jump.

Draw what a product depends on and you get a Wardley map. Then three kinds of people: explorers who try things, villagers who care about the user and package it, town planners who make it reliable and boring. Work out your centre of gravity, and chapters carry badges for each. I read this one properly.

This is the first chapter that feels about me. I think I am a villager: I love the customer being happy. It makes me feel I have a role rather than being the person who cannot code. The map definition is heavy and the table of stages is skippable, but the three people are great.

They are the ones who let things scale; without them, everything would die before it reached the other side. ✓ verbatim

Question: Which badge does a chapter about shipping a booking tool carry for someone like me?

Confusion: The table of what evolves (practice, data, knowledge, backup restore time) lost me; I did not know why it was there.

Next: Good. Worth reading more. Chapter 5 is next.

Steve Jobs on connecting the dots backwards, being multidisciplinary, why maths is taught badly, think with the AI and do not delegate your thinking, then the author's pivots from one product to another. I skimmed the middle and read the first lines and the exercise to write down three influences.

I liked 'if you hated maths at school you were probably taught it badly'. But we are on the fifth chapter, and the author's own products (SG/Send, The Cyber Boardroom, MyFeeds) are all over it. I do not yet know why I should care how his experiments went. I have not been told how to do anything.

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

Question: When do I get to start building something?

Confusion: Names like SG/Send and MyFeeds come with no explanation of what they are.

Next: Five chapters in, no weekend plan. I will look at part 2 and, if it is still not practical, jump ahead.

The frame: what generative AI is, a prompt and an agent loop, why everyone who can brief an agent is a developer, and explorer coding. A short intro.

'Everyone who can brief an agent is a developer' is a chapter title for me. I will go for that one.

Next: Skim chapter 6, then go to chapter 8.

AI is a big bucket; generative AI predicts the next word and is built from an enormous picture of how things connect. Hallucinations are good for stories and bad for facts. Then a long list of corrections the author's agent made to him. I read the first half and skimmed the list of corrections.

The first half was a good plain explanation and I was following it. The 'Where I was corrected' list is for people who already know what weights, diffusion models and tokens are. That is the second hard part in a row for me.

A language model represents words and their contexts as long lists of numbers ✓ verbatim

Question: Is the author saying his own explanation is wrong?

Confusion: The corrections list (vectors, attention, diffusion, emergent abilities) was a pile of jargon and I could not tell which corrections I needed.

Next: I use ChatGPT already so I will not stay for this. Chapter 7 is next but I will skim it.

A prompt is a lens, a role is a model with a description, an agent is a loop with a goal, tools, why a long chat forgets, and keeping one topic per thread. I read the lens part and the sections called 'It remembers nothing' and 'Many focused threads'.

The lens idea is useful and the 'long chats are compressed' bit explains something I have noticed at work, when ChatGPT forgets my early instructions. But harness, tool calls and the 18 summaries of a site called sgit.ai were again someone else's world.

Not magic: a lens, a role, a loop. Understand that, and you are the one who sets the constraints. ✓ verbatim

Question: What is the harness and do I need to understand it?

Confusion: I did not know what 'the program around the model' is in practice.

Next: Okay. I have given it three hard chapters out of seven. Chapter 8 by its title is mine.

If you can brief an agent you are a developer. The spreadsheet is the first piece of software people build themselves, with real horror stories of spreadsheet errors. Portuguese, or any language, is a programming language. Then what the briefing-only developer cannot do: security, maintenance and checking.

This is the chapter I came for. I run campaigns on spreadsheets, so the spreadsheet stories felt real, and 'the language you already speak is the one you program in' is exactly my fear removed. The honest bit about what I cannot do is what stops me feeling stupid and also warns me.

The language you already speak is the one you program in. ✓ verbatim

Question: What would I actually brief the agent to build for my sister, and in what tool?

Confusion: The 'non-functional requirements' list (security, observability, versioned) goes by with no example.

Next: Yes, read on. I want chapter 9 to tell me how to start.

Vibe coding is renamed explorer coding. An explorer app is a brief, not ready for customers. The gap to a real product is a handover to people who maintain it. There is a long story about the newsroom's meter and a villager's craft.

I get that I should not put a half-built thing in front of customers. But this is aimed at people who have developers to hand over to, and I do not. The middle with Linux kernel rules and a Harvard article was not for me.

An explorer-coded app may look like an application and feel like an application, but it is a brief. ✓ verbatim

Question: If I build the first version, who becomes my villager and town planner?

Confusion: The 'bonus test' and the Developer Certificate of Origin paragraph meant nothing to me.

Next: Enough theory. I want to try something this weekend, so I will go to the hands-on chapters.

One afternoon with Google Drive, Gmail and an AI assistant: a folder per person with five documents, a six-line rule list for the agent (draft, never send), five exercises across a week. Then why the agent should not run on your laptop.

This I can do. I work with customers and a messy CRM spreadsheet, so a folder of customers with an assistant that drafts follow-ups is something I would try on Saturday. It does not build my groomers' tool, but it is a real first step, and it never made me feel stupid. The laptop section got more technical but I understood the idea.

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

Question: Is this the same as the thing I would build for dog groomers, or just a way to manage my own customers?

Confusion: 'Claude seat', 'Cowork' and 'Claude Code on the web' appear as if I know the products.

Next: Keep going to the chapters about customers and pricing, which I will jump to.

Four moves: define the product, price it at a number you would be happy to be paid, deliver it at a cost that grows a step at a time, then offer it free to the people who know you. Free is not zero for the customer. Pay after the value, and why pence payments do not work because of card fees.

Now this is marketing and sales, my own world, and I liked 'free is not zero', the card fee arithmetic and not making people apply. But every example is RiskMandate with a vault and a policy, and I cannot picture a pricing ladder for dog groomers from that.

A subscription is a discount for regular use, not rent on something somebody ignores. ✓ verbatim

Question: What would my own ladder look like for a booking tool, and what would the cheap automated tier be?

Confusion: 'An Agent Behaviour Policy for one agent, as an encrypted vault' meant nothing, and I had to guess what the product being priced even was.

Next: Next the hands-on part for a weekend plan. Chapter 30 looks like the start.

The same folder-per-person Contacts workflow as chapter 13: five documents, six lines, five exercises, and a note that the folders can move into a vault later.

I have read this already. It is nearly the same text as chapter 13, with the same six lines and the same five exercises. The tone is shorter so it is useful as a checklist, but it makes the book feel padded and the editor in me notices.

Every mistake adds a line to the policy. ✓ verbatim

Question: Why do I get this twice?

Next: Skip the other hands-on chapters that look like copies. Jump to the price and customers one, then the weekend one.

Price it, offer it free to people who know you, do not make them apply, set a stop rule before sending, and count in a file: created, customised, used. Then take it away and see if they miss it. Repeats chapter 19 closely.

This is a good practical list and the counts (created, customised, used) are clear. But most of it is chapter 19 again, with the same RiskMandate example. It would be a good one-page checklist but I have already read the reasoning.

A thing created and never customised is a download. Customised and never used is a document. ✓ verbatim

Question: Where do I find my first ten people if I do not already have a list?

Next: Last one: the weekend chapter. If that is a how-to for me I buy it.

Getting the job or making it: a part for people going into a job, and the way to show it is acting as the job you want. Short intro.

I have a job and want to build something, so this part is probably not mine. Fine for a graduate.

Next: Skim the three chapters quickly.

The advantage in 2026 is running agentic workflows, with a history of skills that got people hired and then became ordinary. Sections on being more technical than you think and doing it to see it.

As a beta reader: it makes sense for a job hunter but I am not one. It is the same 'brief an agent' idea as chapter 8.

In 2026 I think the answer is one thing: running agentic workflows. ✓ verbatim

Next: Next.

Act as the job you want: manage something for a startup or open source project and publish it. A placeholder section on the job hunt itself.

Good advice for the person who needs it, and honest that one section is a placeholder. Not for me this weekend.

And if you have time, the best way to get a job is to act as the person in the job you want. ✓ verbatim

Question: What happened to the woman at the startup summit?

Confusion: The placeholder 'job hunt' section tells me the chapter is incomplete.

Next: Next.

Publish constantly: the author's own books stall, his articles do not. A book is an ordered set of finished articles. A loop you can copy on day one and speaking instead of typing.

This could be useful for a marketing person like me, an article behind every answer, but the chapter leans on the author's newsroom setup and the 'voice is not a setting' passage was hard to place.

I believe in publishing constantly, and my books stall. ✓ verbatim

Confusion: Does the newsroom assume I use GitHub?

Next: Next.

How to work with an agent that takes everything literally: where it runs, the brief, mistakes becoming rules, one session per job.

Fine. This is the part I would have liked straight after chapter 9.

Next: Chapter 14.

An agent is a session with a focus; the brief is the unit of work; mistakes become rules; lessons files and debriefs. The agents merge and share state.

The first half repeats chapters 7 and 13 about one session per topic. The merge and shared state sections go into how the author's own newsroom agents work.

The most important habit is the simplest one: one session per topic. ✓ verbatim

Question: How does a non-coder keep a lessons file?

Confusion: Briefs and merges between agents assumes I run several agents.

Next: Next part.

Direction and judgement: the scarce skill, who answers for a decision an agent made, and facts and evidence.

Short and relevant to a founder.

Next: Chapter 15.

When agents can build anything, choosing what to build is the scarce skill. Places of highest leverage, knowing when to stop, a budget per step, where a person belongs.

Useful and fairly clear: do the few things that are not blocked and would change a lot. It reads like the author talking to himself in places.

Building stopped being the bottleneck the day anyone could brief an agent. ✓ verbatim

Question: How do I know my idea is a place of highest leverage?

Next: Chapter 16.

Ask for the evidence, provenance, how to disagree rationally, reality corrects the graph, be the best prepared person in the room.

Mixed. 'Be the best prepared person in the room' is a good ending but the middle uses graphs and provenance again, and two notes to Dinis are still open.

One of the most important skills this generation needs is questioning, and asking for the evidence. ✓ verbatim

Confusion: The 'Reality corrects the graph' section lost me.

Next: Part 6, the centre.

Start the company: execution got cheap, ship, price, sell, fund, and make the company real. From experience, not legal or tax advice.

This is what I am here for. I will try to give it more attention.

Next: Chapter 17.

Be a founder, literally. The playbook that worked (raise, hire, build) against experts who now build first. Where the opening is, where the company sits, and a company in all of this.

Encouraging for a first-time founder with domain knowledge, and 'experts were always the better bet' helps. A note says two of the three expert founders are people the author helps.

In 2026 I also mean it literally, because the thing that stopped most people from founding anything has just become cheap. ✓ verbatim

Question: Who are the three expert founders, and are any of them anything like me?

Confusion: I could not tell how a marketing person fits 'expert founder'.

Next: Chapter 18.

Be profitable, let investors come to you, open source everything. A startup is a thing somebody wants to buy at a profitable price; shipping means somebody else can use it. A story of five days, ten customers and the test of whether they miss it.

The shipping definition ('not a demo or video of a demo') is the clearest thing in the book. The SG/Send story is long and repeated elsewhere.

Not to plan, not to pitch, not to architect. Ship. ✓ verbatim

Question: What counts as 'miss it' if my customers are dog groomers with no time?

Next: Chapter 20.

Sell by answering a question already in the room, with a deadline; sell value that lands the day they switch; sell a service before the product; a marketplace of people; and what the buyer is about to do to you.

As a marketer I recognised 'answer the question already in the room'. But the five questions are about contracts, insurance renewals and a board's risk appetite, not a small shop.

It is to answer a question somebody already has, with a deadline, in the language of the room. ✓ verbatim

Question: What is the equivalent question in the room for a dog groomer?

Confusion: The enterprise examples assume a buyer with procurement.

Next: Chapter 21.

Do not raise before you are profitable, who investors choose and why, selling shares instead of issuing them, what breaks. A music analogy.

The record deal analogy works. Shares, tax and survivorship are fairly dense. I would stay with it as a founder but it is not what I need first.

The worst possible time to raise money is before you are profitable, before you have the product, before you understand the fit. ✓ verbatim

Question: Do I need any funding at all for a small booking tool?

Next: Chapter 22.

A placeholder that says so: two lines on Portugal and Lithuania, nothing on the UK or the EU, and a section on hiring with a remote workforce.

This is the chapter that would have told me how to start a UK company, and it is empty. If this were the version I paid for I would be annoyed.

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

Question: Where do I find the steps to form a UK company?

Confusion: The Upwork section does not fit the title.

Next: Part 7.

Build it without a server: serverless, privacy by design, control of agents, open source, and shipped means the user has it.

The word serverless is the first jargon; the intro defines the goal in plain terms (nothing to run at three in the morning).

Confusion: Serverless is not defined here.

Next: Chapter 23.

Ship serverless: nothing to pay when idle. A table of cloud products and whether they cost anything when nobody uses them, then Lego blocks, the file system as database, no server by design.

A table of Lambda, Cloud Run, Cosmos DB and Aurora is a wall for me. The rule at the top is clear and useful, the rest is for a developer.

A founder with no servers has nothing to run at three in the morning and nothing to pay for when nobody is using it. ✓ verbatim

Question: Which of these would a non-coder ever pick?

Confusion: Nearly all of the product names.

Next: Chapter 24.

sgit is git for files you cannot put on GitHub, encrypted before it leaves the machine, with a server that holds only ciphertext. Keys, write-only lanes and two corrections.

This is the author's product, and reading it, I cannot tell whether I need it. It is dense with AES-256-GCM and hashes.

There is no key escrow, no admin override and no support engineer who can recover it for you ✓ verbatim

Question: Do dog groomers' customers' phone numbers need this?

Confusion: AES-256-GCM, SHA-256, multi-parent commits.

Next: Chapter 25.

Write down an agent's reach, its mandate and the gap, with a barrier for each, which is an Agent Behaviour Policy. The policy is the business, who can pull the plug, secrets and small vaults, the model cannot be the barrier.

The idea of reach versus what you agreed is clear and quite sensible for an assistant with my email. But it is also a product pitch for RiskMandate.

A control bounds a grant only if it is enforced by something the grant does not include. ✓ verbatim

Question: Does this apply to the folder assistant in chapter 13?

Confusion: The four kinds of barrier needed one more example.

Next: Chapter 26.

Why to open source: nobody can lock you in. What open source is and is not, forks, not free, all of it or it is an evaluation version, sell the build not the code, and your first pull request.

I did not know what a repository was and this is the first chapter to say what a licence is. It is clear enough but it needs a note on what open source means to someone who has never opened GitHub.

It was never about contributions. It is that nobody can lock you in, and that includes you. ✓ verbatim

Question: Do I have to open source my booking tool?

Confusion: 'Fork', 'pull request' and 'repository' are used before they are explained.

Next: Chapter 27.

Two releases of the author's site pushed cleanly and never reached anyone for 40 minutes. Green does not mean live, tests that never run are not tests, run it against a twin, five people who do not exist.

A good story about 'shipped means the user has it', but it is about CI deploy jobs and nginx configs, and I could not follow the cause.

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

Question: How would I tell my sister's customers can actually get to my page?

Confusion: Release script, deploy job and third-party action.

Next: Part 8.

What lasts: graphs as memory for your future self, and civic responsibility.

Soft for a founder weekend.

Next: Chapter 28.

Backup your life, graphs as the way the author thinks, memory is context management, publish and the site becomes the memory.

Backup your life is common sense. The graph material repeats chapter 3.

You basically need to backup your life! ✓ verbatim

Next: Chapter 29.

The author's keynote in Lisbon in 2010 and what came true in Brussels: product liability, the cyber resilience act. State plumbing, commons and sovereignty is the ability to move.

This is the author's country and policy history. I have no stake in it. It would cost the book readers like me.

Code is law. You are one of the most important pieces of the puzzle. ✓ verbatim

Question: Is a Portuguese edition coming and is this chapter for it?

Confusion: EU directives and the commons.

Next: Part 9.

The hands-on part, dated 2026, with eleven workflows from an agentic workflow from zero to a serverless product in a weekend.

This is the part I want, and it comes after about 29 chapters of reasons. I would move it up.

Next: Chapter 31.

Draw a graph of something you know: a text file of about forty lines, an answer to a question people keep getting wrong, and a list of things nobody knows. Seven steps, why the rules are what they are, the first week and the agent.

The advice to pick a question, not a subject, is good and short. It is a text file and a seven-step recipe I could follow, once I buy the idea of a graph.

The most common way this fails is picking a subject instead of a question. ✓ verbatim

Question: What is the first question for a dog groomers' tool?

Confusion: Nodes and edges again, with a reference to a graph book I have not read.

Next: Chapter 32.

Pick one idea and have a chat tool explain it five ways: for a teenager, a musician, a doctor, a football fan, and in another language. Then ask to be corrected.

An easy hour and a good exercise for a Z reader. It repeats the lens section of chapter 7 almost word for word.

The lens was wrong, not the person. ✓ verbatim

Next: Chapter 33.

Run an agent loop by hand with yourself as the harness: write a role file, give it one tool in words, loop, and see why it forgets.

I can see this working as a no-code exercise and it would make chapter 7 land. Good.

The model never runs the tool. It writes a structured request, a tool name and its arguments ✓ verbatim

Question: Would I do this before or after chapter 13?

Next: Chapter 34.

How the book's storyboards are made: cast file, data per story, drawn in the browser, quotes checked. Poses, moods and Rough.js.

This is about how the author draws his own storyboards. Not mine. A cute idea for a team that makes decks.

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

Question: Do I ever need this to start a business?

Confusion: Rough.js, a seed per panel and exports.

Next: Chapter 35.

Three habits around every meeting: brief yourself, brief on the people, follow up. A short chapter that says the newsroom has not written the workflow up yet.

For an account manager this is a solid idea and I do it already by hand. But the chapter is a shape waiting for the author's steps.

Now it is a brief and a feedback loop. ✓ verbatim

Question: What are the author's own steps?

Confusion: The notes to Dinis show it is unfinished.

Next: Chapter 36.

Set up a public GitHub site written from voice memos and an agent, and an inbox worked by an agent that drafts and never sends. Eight steps.

Public GitHub repository and Actions and Pages are all new to me, but eight numbered steps are the form I want. Needs a first-timer's walkthrough.

The agent writes the code. You do not. ✓ verbatim

Question: What happens if my repo is public and I put something private in it?

Confusion: GitHub Pages, Actions and a validator.

Next: Chapter 37.

A chapter on forming your company that says it is not written yet: what is known, and what to decide before you register.

The promise on the cover is a company with a bank account by the end, and this is not written. I would be upset to buy this.

By the end of this chapter you will have a company: registered, with a name, a country and a bank account. ✓ verbatim

Question: Is there any date for the UK steps?

Confusion: The chapter opens with a promise and then says it cannot keep it.

Next: Chapter 38.

A table of cloud and agent startup programmes, as of 7 October 2026, with amounts and eligibility, and a rule: apply with the work you have published.

Useful to someone with a company, with a good honest warning that it dates. The programmes are developer-infrastructure ones, not marketing or small shops.

Programmes change; check before applying. That sentence is part of the table, not a disclaimer ✓ verbatim

Question: Which of these would take a solo non-technical founder?

Confusion: E2B, Daytona and Scaleway meant nothing to me.

Next: Part 10.

How the book was made, and what you made: the repository, the vault, roles, articles, decisions, release candidates and the measure of companies created.

Process again. I would put this at the back or on the website.

Next: Chapter 41.

How the second edition was made: repository versus vault, roles, articles as briefs, storyline, graph, release mode and board, co-written with Claude and credited, what you are reading, and what you made.

The honest credit to Claude is nice and fits the book. The long process description is for people building a newsroom, and two of the author's open questions are still in it.

One rule organises everything. The repository holds the official versions; the vault holds the working versions. ✓ verbatim

Question: How do I tell the author what I made, and is the list public?

Confusion: Roles like cutter, advocate and librarian.

Next: End.