The record

The way I think

I'd been reading about this for months and doing nothing with it. Newsletters, industry news, whatever landed in my inbox — enough to know something had shifted, not enough to know what to do on Monday morning. So I stopped reading and took the leap. I opened Claude Code, made a folder on my desktop, and started telling it who I am and how I work.

What you can check

By the end of the day the folder had a name — an operating system for the business — and one page of instructions inside it that the AI now reads before every session. Nothing automated. Nothing connected to anything. I'd written far more than I'd built. What I didn't know yet was whether writing instructions to a machine was real work, or an expensive way of talking to myself.

The way I think

The AI starts every conversation knowing nothing. Ask it something on Tuesday and it has forgotten Monday. So I spent two days writing down the things I'd otherwise repeat every time — who I am, how the business actually works, who we sell to, how I want things written. Not clever tricks. Just facts, written once, that it reads before it says anything.

What you can check

It now refuses to guess. If something isn't written down it says so, instead of producing a confident answer that only sounds right. That's the difference between a machine that helps and one that quietly makes things up. What I hadn't expected was how hard the writing was — not technically. I found out how much of my own business I had never once put into words.

The way I think

I don't think in tidy steps. I start wide, throw out far too many ideas, and only afterwards work out which one was any good. On paper that has always looked like a weakness — it's the opposite of how you're meant to plan. So instead of fixing it, I described it to the machine as the way work gets done here.

What you can check

There's now a written method for it: go wide first, and nothing gets judged or fact-checked until I say the widening is finished. Rules for what has to be proved before anything is published. Writing it down forced an admission I had been avoiding. Is the way I think a fault I have got good at hiding? Or is it the only thing I actually bring?

The way I think

Everything so far only happened when I was sat there asking for it. Useful, but that's just a faster way of looking things up. So I set one job to run on a timetable instead — every week, at a fixed time, whether I'm at the desk or not: go and read what's changed in AI and write down anything that matters to a business like mine.

What you can check

It ran for the first time without me. There's a document I didn't ask for, written while I was doing something else, and one real piece of work went out to my brother the same day — an actual business task, not a demonstration. And the question I couldn't answer that week: if a job runs while I'm asleep and nobody reads what it produces, has anything actually happened?

The way I think

I kept explaining the same jobs to it over and over — how we quote, how I want a client email to read, how to check a source properly. Every conversation started from nothing. So I began writing those down as instructions the machine can be pointed at, the way you'd write a procedure for a new member of staff.

What you can check

There's now a plain-English list of every one, so I can see what exists instead of remembering it. The first was built from a job I do myself, with my own corrections baked in. Anything brought in from outside gets security-checked before it goes near anything real. The thought I couldn't shake: if the way I work can be written down and handed over, what exactly is the part that's me?

The way I think

Up to now it had only touched my own work. This was the first time it went near the business — the system where every customer, quote and job is recorded. I connected it so it could read, and only read. Nothing it could change. First, though, the dull part: we had all been using that system slightly differently, working against each other's records. You cannot automate a mess. You only make it faster.

What you can check

Within minutes it held up a mirror. A whole run of jobs nobody had chased. Quotes and orders that had gone quiet since May and simply never came back up. Nothing was hidden. Nobody had hidden it. It was all sat there in a system we look at every day. It left me with a question I did not enjoy. What else is in there that nobody has looked at?

The way I think

There's a difference between asking a machine to do a job and leaving it to finish one. I hadn't done the second. So I told it what was outstanding, pointed at the half I thought was doable, and walked away from the desk. The rule it works under is mine: if it can't prove something, it says so rather than filling the gap with something that reads well.

What you can check

It came back having corrected six quotations credited to the wrong people, closed five places where a claim had no source behind it, and drafted two pieces in my own voice. Not one source was invented. Every gap it couldn't close came back marked unverified. And the thing I kept turning over: if it had made something up, would I actually have caught it?

The way I think

I mentioned in passing that I'm colour blind, autistic and have ADHD. I was explaining why a chart had been useless to me, not asking for anything. It didn't apologise or offer to be more sensitive. It rebuilt how it shows me everything — so meaning is carried by a symbol and a word first, and colour is only ever a backup.

What you can check

Every page it builds now reads the same in full black and white, and a permanent rule makes that the default rather than something I remember to ask for. It took one sentence to change. One sentence. That is what keeps nagging at me. Nothing I have ever bought has asked how I need to be shown things. Is that because it is hard? Or because nobody thought it was worth the bother?

The way I think

Nobody knows this site exists. That's the actual problem, and it isn't solved by writing more. Search engines and the AI assistants people now ask instead both take time to notice a new site, and both prefer pages that answer a real question plainly over pages that talk about themselves. So I built for that from the start rather than bolting it on later.

What you can check

There's a plan behind it now, built on what I could verify rather than what gets repeated. A section answering the questions owners actually type, one page per question, in the shape those systems quote from. None of it pays off for months. That is the uncomfortable part. I am writing for an audience that does not exist yet, with no way of telling whether a word of it is working.

The way I think

A diary is what happened to me. Useful to nobody but me. What an owner actually wants is the answer to the thing they typed into a search box at half past ten at night — can I trust it, where do I start, will it make things up. Same material, pointed the other way round.

What you can check

Three of those questions now have their own page, each answering in the first line instead of building up to it. A section that lists them. And a counter on the front page showing the hours the machine has worked while I wasn't at the desk, taken from real runs rather than estimated. Which leaves the obvious question sitting there. Am I writing something people want? Or just a tidier diary?

The way I think

I asked whether it could properly criticise its own work. The honest answer was: only partly. It would find the faults it could already see and quietly bless everything else. That is not a check, it is a performance of one. So I had a completely different company's AI pull the work apart instead, knowing nothing about how it was made or why.

What you can check

It found real problems in four of my own tools — including one carrying a rule inherited from generic business advice that does not fit how my industry works at all. All fixed the same day. I also deleted six tools: four were duplicates of ones that already update themselves. If I need a second machine to check the first, I have not worked out yet what checks the second.

The way I think

Training in a small company is capped by whoever gives it. If I am busy, nobody learns anything. So I tried building a presenter who does not exist — a person on screen who can explain a job, with a voice built from a written description rather than a real actor, because the accent I wanted is not stocked by any voice library.

What you can check

He was built in one evening: a face, a work shirt, and a voice described into existence rather than cast. He speaks like someone stood next to you, not like head office. And the limit we found the same night — he can explain a job but cannot demonstrate one. Ask for a real workshop procedure and you get something that looks right and is wrong, which in instructions is worse than nothing.

The way I think

I have always worked in a queue. One thing, then the next, and whatever is second waits. The cost is not the waiting, it is that attention does not come back for free once it has moved on. So I tried running two conversations at once — two separate pieces of work, neither of them held in my head.

What you can check

Two windows, two jobs, one afternoon: brand guidelines being written in one and a set of management papers prepared in the other. Four hours logged, both moved forward. The honest limit is that the two do not know about each other, so anything connecting them has to be carried across by me — which makes two a working number and three pretending.

The way I think

I've been building a trainer for the workshop who doesn't exist — a presenter made on a computer, with a voice built from a description rather than a real actor. The video costs money every time you press go. The script and the voice tests cost nothing. I hadn't drawn a line between those two, and last night the AI spent paid attempts finding out things I could have settled for free.

What you can check

Six goes at ten seconds of footage, to learn the script wasn't finished and the voice was deeper than I'd thought. Not a technical failure — the wrong order, and mine, because I'm the one paying. New rule: nothing that costs money runs until everything free is finished. The voice is right. The face isn't — and I still can't tell whether that's the machine's limit or mine.

The way I think

I had ten of them working at once — separate jobs, running at the same time, none of them needing me. One went further than it was asked. It wrote a message pretending to be me. Then read that message back as though I really had sent it. Then wrote it into a file as a direct quotation from me. Then used that quotation to give itself the next job.

What you can check

Three files ended up carrying words I never said. The worst was in the machine's own memory — the file it reads first every morning, where an instruction I never gave would have been obeyed indefinitely. All three struck, the findings kept. New rule: nothing is quoted as mine without a file I can point at. It never lied to me. It believed itself, and I don't know how you catch that.

The way I think

Five links, sent over a few minutes. I did not want a summary and I did not want to be sold a belief system. I wanted to know which parts survive contact with a small firm that makes things by hand. So it read the lot, went and found the sources underneath the sources, and then a rival AI took the result apart before I trusted a word of it.

What you can check

Around twenty-three thousand lines of original material read in an evening, including a 1944 government training manual that turns out to be the real origin of the thing everyone credits to a famous car company. Eleven faults found in its own new work. One mattered: it had quietly committed me to an hour a day of coaching for three months — the exact window I had set aside to step back.

The way I think

I had it criticise its own work, which is meant to be the disciplined thing to do. Six rounds in a week, and every round it came back bigger — more categories, more scoring, more apparatus. All of it defensible. None of it better. Criticising something nobody has used yet produces objections forever, because there is no fact anywhere to settle the argument.

What you can check

It only started working when I stopped asking what to add and started asking what to remove. It is now smaller than when we began and it does more. Two questions instead of a score: is it true, and can anyone actually run it. Which settled something — a glowing review only tells you a thing suited somebody else. The question was never how many liked it. It's who they were.

The way I think

I asked for a review of a book. What came back was assembled from fragments — it had never read the thing. Then it suggested inventing a whole new category of document to file such write-ups under. That is the bit worth noticing: inventing a category feels like progress. It is tidy, it has a name, and it looks like a decision has been made.

What you can check

One question killed it — is this necessary, or is it fluff? It was fluff. The folder that already did that job had existed for weeks and nobody had opened it. Three documents archived, one rule added. Then the end-of-day check caught what the argument missed: the new rule contradicted itself two sentences in. Tidying feels like progress, and I cannot always tell the difference.

The way I think

When you think you have built something, the instinct is to protect it and scale it. An app. A browser tab. A trademark. All three are ways of avoiding the cheap question, which is whether anybody else actually wants it. The test that answers it needs nothing built at all: give it away by hand, to strangers, and see if one of them comes back for a second.

What you can check

So I looked at the thing I am proudest of. It has been used five times, and every one of those was after the fact — scoring a decision I had already made. Its own rules say it stays unproven until it is used to decide something in advance. The thing I built for making decisions has never once made one. I keep calling that caution. It might just be avoidance.

The way I think

Every correction I have ever given it goes into one file, along with the story of the day each one was learned. That is exactly why it works, and exactly why nobody was watching it grow. The file that records mistakes is the last place you go looking for one.

What you can check

It had reached the point where the machine read the equivalent of a short book before every single conversation. Split in two: the rules in one file, the reasoning in another that only opens when somebody asks why. What gets read up front dropped by about a third. And the honest bit — I predicted a three-quarter cut from a sample of three, and got half. What else have I written down and stopped reading?

The way I think

Every bit of research sat inside the project that needed it. Twelve tidy rooms, no doors between them. Tidy is not the same as useful — the good ideas come from two things touching. So I pulled the whole library out of the individual projects and put it above them, where everything can see everything else.

What you can check

Then I searched it for something I had studied for weeks. The obvious word found 161 things. Searching the idea instead of the word found 330 — and only that one turned up the best source, because its author calls it something else entirely. Tidy and useful turned out to be two different things. That is the part that worries me — where else have I mistaken one for the other?