The Technical Account

I am not a "writer" though I write a lot down. With an AI I wrote a book.

Both halves of that are true and I've stopped being embarrassed by either. The first one I said so often during the project that it became a running joke, usually at the moment I was about to ask for something difficult. The second one is a sixty-five-thousand-word memoir that exists, that I'd put my name on, and that I could not have produced on my own in any amount of time.

This is the account of how that happened. It is as much about what went wrong as about what went right, and before you read a word of it I want to explain why that is, because it isn't confession and it isn't a warning about the technology.

What the other book is, and why you don't need it

There's a memoir. It covers a life that went from teaching windsurfing, to a stretch on television, to running digital transformation for a large company, to captaining yachts across oceans, to building software I had no business being able to build. Along the way there's a broken ankle that went badly wrong, some relationships I handled poorly, a family fight over an inheritance, and a hair transplant in Turkey.

You do not need to have read it. This book stands on its own and when it refers to that one it will tell you what you need in the sentence where it comes up. They're different books for different people. That one is for anyone who recognises their own patterns in somebody else's. This one is for anyone about to attempt serious writing with a machine.

Thirty years of scrawls

The thing that made this possible is not the AI. It's what was already in the cupboard.

I have always written things down. Not diaries, nothing so organised. A biography I wrote in 2009 for a proposal that went nowhere. Another attempt in 2017 when I was trying to make sense of a bad year. Notes on my phone at two in the morning in the middle of an ocean. Emails I sent that said more than I realised. Accounts of things typed out afterwards because something needed saying and there was nobody to say it to.

When this started I went looking:

> I have also been through mac notes and saved out any scrawls I made over the years to the project folder as txt files.

That's the raw material. Thirty years of it, kept without a plan, because I've always written to hold onto things rather than to make anything of them.

None of that came from a machine and none of it could have. Which means the honest version of what happened here is narrower than the headline suggests. The AI did not give me a life, a memory or a judgement about what any of it meant. It supplied labour and craft, at a scale and a speed I could never have reached alone, over material that already existed. It made connections I would never have seen reading back scraps of notes.

That is still enormous. It is also not the same as writing my book for me, and the difference is most of what this book is about.

The tool

It was Claude, mostly in long working sessions with the files in front of it. The model changed over the six months, as they do. I'm naming it because a book about doing this that won't say what it was done with is no use to anybody, and because several of the faults in here are specific enough that knowing the tool matters.

None of this was possible in an upload, download chat window with a tone of voice prompt, and it was not possible on a basic subscription plan either. At times the entire book had to be held in memory at once, alongside the supporting files that map the threads, track the contradictions, carry the rules and hold the tone of voice to be read before anything else. And that list kept growing.

Nothing in this book is a complaint about it. It is far better at writing than I am. That turns out to be the problem, and it takes a while to explain, which is what Part Two is for.

Why I'm telling you the bad parts

A few years ago I had a hair transplant. The plan was to have it done and recover quietly at sea, which would have been easy, because nobody sees you out there.

Instead I posted the whole thing publicly. Scabs and all. Turned up at the gym with a head that made it obvious what I'd had done, in the period where it looks its worst.

I did not expect what followed. Men I barely knew got in touch wanting to talk about it. Friends went and had it done themselves, having said nothing for years. And what made the difference was not that I'd had it done. It was that I'd shown the ugly bit, which nobody does, so everyone assumes the ugly bit is a reason not to.

That's what started the memoir. Showing the ugly bit worked for a head of hair, and I wanted to do the same thing on a much larger scale:

> Its this on a much larger scale that is motivating me right now.

And underneath it:

> If I can help just one person figure it out faster than it's taken me, then it's a worthwhile process.

This book runs on the same principle. It would have been easy to write it as a how-to, with the method laid out and the bad parts left out. That version would be tidier, and it would leave out the part you most need.

So this is the scabs-and-all version of writing a book with a machine. Not the polished account, not the case study where everything works. The four months where an instruction I wrote myself was quietly dissolving my own sentences. The week I spent counting punctuation that turned out to include two headings. The chapter that scored highest and was the worst thing in the book. The fortnight the project spent going backwards while every individual thing we did was correct.

You could write a version of this with all that left out. It would be shorter, it would be more impressive, and it would be no use to you whatsoever.

What it costs, which is a separate thing

There's a line I wrote down partway through, about what this kind of work asks of you:

> For this to work, I have to listen to my own voice, see my own face, suck it up! And put out there what's usually locked up.

That's about the memoir rather than this book, but it holds for both. Writing about your own life with a machine means reading your own sentences back more times than is comfortable, having your own phrasing handed to you improved, and deciding, over and over, whether the better version is still you.

And it means being the only person in the room who can tell when something is wrong, which is a specific kind of tiring that nobody warns you about. The AI will go anywhere with complete confidence and it will not tell you when it's lost. My development work taught me that first. The book taught it to me again.

What you'll get

Four things, and then the honest one.

What held for six months and never broke, which is the part worth copying and the shortest chapter in the book.

What went wrong invisibly, where the output was good and wrong at the same time, and none of the checks could see it.

What went wrong while everything was going right, which is the part I had never heard anybody describe, and which nearly ended the project.

What only a person can do, which is less than I'd hoped and more important than I expected.

And then the honest one, which is that almost every guard in here exists because something broke first. None of it was designed. It's scar tissue, and it's transferable even though the lessons weren't learnable in advance.

I can't save you the six months. I can give you the guards it took me six months to build, so that you spend yours finding the next set.

In June we went looking through my whole workspace for a specific kind of lie.

The question was this. Of all the files sitting in there labelled as my writing, my actual typed words, how many of them were secretly the machine's? Because if even a few were, the ground the book stood on was rotten. Every chapter was being built by pulling quotes out of those files. If the files were contaminated I'd have been quoting a machine back to myself and calling it my life.

The answer came back: none. Not one file was a write-up wearing my name. The labelling had held.

I want that on the record before anything else. This chapter is about what went right. Underneath everything that broke there was a spine that never gave, and it's the part worth copying. If you take one thing out of this book, take this one.

Here's what the spine was.

Three kinds of file, and never mix them up

Everything in the workspace is one of three things.

There's stuff I typed. The scrawls from Chapter 1, plus my accounts and notes, typos left in. That's the ground floor.

There's chat transcripts. Half mine, half the machine's. My lines in there are as good as anything I typed. The machine's questions and summaries in the same file are not mine and must never be quoted as if they were.

And there's everything the machine wrote about my material. Indexes, pattern files, session notes, summaries. Useful for finding things. Not evidence of anything.

Then one rule that follows from it: when two of them disagree, the thing I typed wins. Always. And if you're not sure, use my exact words in quotes instead of a tidier version of them.

That sounds obvious written down. It is the single most valuable thing in the project. It's what made everything else checkable. Whenever something went wrong later, the three classes were what we checked it against, and that is the only reason the problems could be found at all.

A register of names, checked before you type one

There's a file that lists every recurring person in my life, what they are to me, and which file each fact about them came from. You check it before you put a name in anything. You don't delete from it without asking me.

Sounds like admin. Here's why it isn't. I have a grandad Roger and my dad's brother is a Roger and my other grandad is Ralf, and my dad drove the lorries while grandad Roger ran the office of the same scrap yard. There was a question about who brought the windsurfing kit to the bungalow in Norfolk, grandad or my uncle. That got asked, answered and written down once, and after that nobody had to guess.

With the register, that question was answered once and stayed answered. Without it, a wrong name reads perfectly well, and your grandad is now saying something your uncle said, in a book your family will read.

Anything that can go wrong silently needs a register like this. Names were the obvious one. By the end I'd learned the same trick applies to decisions, which is a later chapter and a harder lesson.

Start a fresh chat, and know why

Two files in my transcripts folder tell this story between them without any help.

One is called "Research book narrative and core story, chat compressed, lost beginning." The other is called "Book writing session fresh start."

The first one is a conversation that got too long. When that happens the thing squeezes what came before into a summary of itself and carries on, and it carries on beautifully, and the beginning is gone. Not corrupted. Gone. The opening of that interview, where I first talked about my family, only exists as a description of itself.

The second file is what we did about it, and it opens with me saying so:

> We decided in another conversation to start afresh to avoid compacting.

After that, a new chat per chapter. The standing instructions get pasted in at the top every time, the source files get read fresh every time, and nothing important is left sitting in a conversation's memory where it can be quietly compressed into a summary.

It feels wasteful. Twenty-five chapters, twenty-five cold starts, the same setup pasted every single time. Do it anyway. The alternative is a conversation that sounds like it remembers and doesn't.

Tell it what belongs elsewhere

Every chapter brief carried a list headed "held for later, do not pull forward."

It named the material that turns up in that chapter's sources but belongs to a different one, which one that is, and why moving it early would hurt. A moment that belongs somewhere much later gets a line saying where and why. A second person who shares a setting with the first gets a line saying they are a different story with a different point.

Without that list, a chapter will happily eat the best bits of the ones after it, because it has no idea they exist. It only sees the pile of source material in front of it and the instruction to make this chapter good. The list is why each chapter kept its own material, and why the later ones arrived full.

Write down what you did, at the foot of the thing you did it to

Every chapter has a note underneath it recording what changed, what was decided and what is still open.

Seventy-two session logs, one for each working session, saying what was discussed and what was still hanging.

This is the practice I'd defend hardest. The notes were right and writing them was right. The one thing to know now is that writing something down is only half a system, because something has to read it back. That comes later in this book. The answer was never to stop writing the notes. It was to make them gettable.

And the part that was thirty years old before any of this

None of the above would matter if the cupboard was empty.

I described what was in mine at the start of this book, and the only thing worth adding here is that I had not kept any of it on purpose. There was no plan. It was thirty years of writing things down and not throwing them away, and it turned out to be the foundation everything else in this chapter sits on.

Which makes it the one item on the list you cannot go and set up this afternoon. Everything else in this chapter takes a morning. This one takes decades, or it takes starting now and being patient, and there is no third option. If you are thinking about doing this at some point, the useful thing to do today is not to plan the book. It is to put somewhere safe the things you have already written without meaning anything by them.

What it adds up to

None of the six things above failed. Not once, across six months and four hundred and thirty drafts. Everything that went wrong, and plenty did, went wrong inside them.

That's the useful bit. Because these held, everything that did go wrong could be found, named and traced. A method that fails everywhere teaches you nothing. One that holds nearly everywhere shows you exactly where the weak points are.

So start here. Three classes of file. A register for anything that corrupts silently. Fresh chat, cold start, every time. Say what belongs elsewhere. Write down what you did. And have something of your own to work from, because that's the only part nothing can do for you.

Get those in place and you've got a workshop. What happens in it next is the rest of the book.

I did this the wrong way round.

I spent six months finding things out the hard way, and then, near the end, I went and looked to see whether it had already been worked out. It had. Most of it. Some of it before I was born.

I could have left that out. Nobody would have known, and the book would have read as a man discovering things nobody had noticed before. I'm putting it here instead, at the front, before you've read a single thing that went wrong, because it changes what the rest of this is for.

The things I thought I'd found

The chapter after this one is about a man being turned into somebody who barely spoke, by a process where his description passed through a summary on its way to the page. I thought that was a thing about machines.

It was done in Cambridge in 1932. A man called Bartlett gave people an unfamiliar story and had them retell it, then had the next person retell that, down a line. What came out the other end was shorter, tidier, and more like the sort of story the people telling it already knew. He had names for the parts of it. Things drop out, the few that survive get exaggerated, and the rest gets bent into a shape that feels familiar. A specific person becomes a type. That is what happened to my coach, ninety years later, with a machine standing in for the students.

I'm told the famous version of that experiment has been overstated for years. The finding holds. The legend around it is bigger than the evidence.

Then there's the scorecard, which I built to tell me whether the book was any good, and which ended up quietly deciding what parts of the book mattered. That is Goodhart's law, from 1975, and there is a sharper version of it from a woman called Strathern that most people quote without knowing whose it is: when a measure becomes a target, it stops being a good measure. There is an even closer fit in the accounting world, where they have studied people swapping the measurement for the thing being measured, and found it happens with no bonus attached at all. Knowing something is counted is enough on its own. That is exactly what happened to me and I had no money riding on it either.

The blind reviews that kept leaking, because the previous scores were sitting in a file every new session opened. Clinical trials sorted that out decades ago and they are blunt about it. A blind that holds only while nobody peeks is not a blind at all. Their answer is to take the thing away, so that peeking is not on offer.

The register of decisions, the one I say is the cheapest thing in this book. Software teams have been keeping those since about 2011 and they call them decision records. And their complaint is my complaint, word for word: people write them and nobody reads them back.

Sealing my predictions in a file before the measurement ran, which felt at the time like an unusually honest thing to be doing. That is pre-registration. It came out of drug trials and it went through psychology after a lot of published findings turned out not to hold up. When you make people say in advance what they expect, far fewer studies come back saying exactly what the researcher hoped. I reinvented it in an afternoon and felt clever.

Making the check a step you cannot skip, rather than a rule you have to remember. Manufacturing has a name for that, and so does design. The research there is not the clean win I'd like it to be. Some of the largest studies of safety checklists found no benefit at all, and the reason seems to be that people perform them without reading them. A step you cannot skip can still become a thing you do with your eyes shut.

And the one I liked least. The bit where I say the writing gets faster and the reading doesn't, so more and more of my time goes on the only job the machine can't do. That's Bainbridge, 1983, on the irony of automation. Automate the parts that are easy to automate, and what the person is left holding is the watching, which we are poor at, while the skill they would need in order to step in wastes away from lack of use. Forty-three years ago. In a paper about power stations.

Even the stupid one has a name. The counter that ran off the end of the last item and told me most of my own archive was contaminated. That is a fencepost error, and it is older than computers. It's what happens when you count the fences and forget there's a post at both ends.

What I make of that

The first was annoyance. Six months. I could have read a book.

The second was relief, and it was bigger. If a man in 1932 hit the same wall I hit in 2026, then the wall is not about me and it is not about this technology. It's about what happens when information passes through something that has to compress it. The machine is new. The shape of the problem is old, and it is going to be there whatever you are using in five years when the tool I used looks quaint.

The third one is the argument of this whole book.

If the problems are old, the fixes are old too. Every one of those fields worked out what to do. Not perfectly, and not without arguments, and I'll point out where they disagree with themselves. But none of them concluded that the answer was to stop. They built the guard and carried on.

That is what I want you to take out of the next couple of hundred pages. This is not a book warning you off. It's a book saying that the things which go wrong here are the same things that go wrong in surgery and software and drug trials and aviation, and that people cleverer than me have already had a go at each of them, and that you can start from their answer instead of mine.

Where they say I'm wrong

There are places the research doesn't back me, and I've changed the book rather than argue with it.

The largest sits in a chapter that used to argue that anybody who grabs a rule is dodging a decision, and that the rule comes out worse for it. There is a mountain of evidence that mechanical rules beat expert judgement, across an enormous range of work, and I'm not taking that on with one memoir. That chapter now says what I did and why, and admits what I can't show.

There's another where I said no instrument could ever tell you when to stop. There are methods for estimating how much you haven't found yet, and I didn't know about them, and one of them I could have run on this book with what I already had.

And I've taken out the line where I said my own frustration was the best detector in the project, because the false alarms were never written down anywhere, so only one side of that competition was ever scored.

The one I can't answer

If I'd done this reading first, would I have written a different book? Would I have found the known method, followed it, and had nothing much to tell anybody?

I don't know. I can see it going both ways. The version of me that read Bartlett in March saves himself a fortnight and a bad afternoon. He also might have trusted the method over what was in front of him, and I'd have ended up with a technically sound book about somebody who wasn't me.

What I can tell you is what happened, which is all this book was ever going to be. One person, one book, six months, no training in any of this. Where somebody has done the work properly, I'll point at them and get out of the way. Where I'm only telling you what happened to me, I'll say so.

And if something in here bumps against something you already know from your own line of work, back your own.