Every quote in this chapter is something I wrote myself, in Spanish. They’re translated here with the typos left in: the typos are half the evidence.
On 8 April 2025, at 14:39, I wrote this:
Sugest names for an AIMA agency I want to start. I’ve thought of AiMAMA but I don’t want to end up being a smartarse.
Six minutes later I was already asking for the logo and a proposal for the company values.
I didn’t have a single client. I hadn’t sold anything or delivered anything, so I didn’t yet know whether that was going to be a company, a weekend side thing or an expensive whim.
But I had the name at 14:39 and the logo at 14:45. And the name that ended up sticking, two years later, was the one I threw out that afternoon for being a smartarse.
The spring of the eight ideas
The previous chapter ended with a year and a half of silence and a weekend staring at a door. This part starts exactly where that didn’t work out.
What I did in March and April 2025 is in the archive and I can’t dress it up: a website proposal for an olive oil brand, suppliers of personalised products, print-on-demand children’s books, generated stories, a flow to issue invoices by itself, a tax assistant, AI tools for financial advisers, automating an Etsy shop.
Eight different things in eight weeks. None finished.
A stranger put it better than I could in a forum, talking about other people: «they built-it, made a singular post to their 12 followers on Twitter/Instragram/whatever, and nobody came».
I wasn’t even getting as far as posting.
The prettiest deck I’ve ever made
In May I finally aimed at something serious, and it wasn’t my idea alone. A friend runs a communications agency and we’d made a split that on paper was perfect: he knows how to sell, I know how to build the tool. We were going to go together to a bunkering multinational, a company in the port where we both had good contacts. My part was the deck.
My own industry, people who spoke my language, processes I knew from the inside. And I threw myself at it.
On 8 May, at 18:17, I asked for this image:
Generate the image of a bot with communication headphones sitting at the computer next to a trader. The image should evoke that the bot is one more person in the office.
Then the client’s corporate colours in the header. Then a cover. Then the angle, which was right and which I still believe: that an agent is one more colleague, taking the repetitive tasks so that people can decide.
After that came the indicators to measure the pilot, a step-by-step guide to the whole project, a script for the first meeting and a tool comparison in case the client didn’t want to leave Microsoft.
I was at it until half past one in the morning.
And at 00:56 that same night, in the middle of all that, I wrote the question that gives me away:
What does Demo version available under NDA mean
I was polishing a document that contained a sentence whose meaning I didn’t know.
I’m not telling this to be funny. I’m telling it because that is exactly what that year was: on the outside, an impeccable deck with its imagery, its indicators and its schedule. Underneath, a bloke learning the vocabulary of the trade at the same time as practising it.
No contract came out of that proposal. Nothing came out of it, in fact: I passed it to my friend to read and I think he never read it. It was never delivered to anyone. Afterwards I tried on my own, with my own contacts, and it’s still there: they’re moderately open to it, but it’s one of those things you have to force through and keep pushing. And I didn’t keep pushing, for a reason I’ll get to shortly.
Something else did come out, though, and it took me a year to learn how to value it.
Four days later, the thing that did work
On 12 May, at 23:00, I started something else. And this one came through the same door: the agency had won a project and had nobody to put on it, because they didn’t have the specialist. The specialist was me, even though I wasn’t one yet. I said yes before knowing whether I’d know how to do it.
That night I put together the project’s objective, an analysis of why it could fail before it even started, the breakdown of the work, a schedule, the map of who decides what, and the introduction emails one by one depending on who they were going to.
It was for a public body in the Canary Islands. And that one did get somewhere: it ended up as a chatbot delivered and running, the first real client I’ve ever had.
The difference between the two things wasn’t the effort, because I put in the same. The difference was that on the other end there was someone waiting for something specific.
And that project isn’t the end of this chapter: it’s the next one. Everything I know how to do today came out of there. Here I’ll just note that it started on a night in May, four days after the deck that went nowhere.
The three things that don’t appear in any business plan
When I tell this year in short, it sounds like I lacked ambition or got distracted. It’s simpler and more uncomfortable: there were three things working against me, and none of the three is fixed with another deck.
The first is that I get paid at the end of the month. I’ve had a permanent, well-paid job at a multinational since 2018. That pays for three children’s school and the mortgage, and it takes out of the way the one thing that forces you to finish: need. Nobody closes a sale at half eleven at night when the month’s bills are already covered. You can want to, and I did want to. But wanting and needing don’t squeeze the same.
The second is that, literally, I couldn’t invoice. On the morning of 19 May I wrote a question that wasn’t about AI or projects:
how can i invoice services to a third party without being self employed and working as an employee?
Twenty minutes later I was comparing going self-employed with setting up a company, and Catalonia with the Canaries. I had decks, indicators, schedules, a logo and a company name. What I didn’t have was a way to issue an invoice.
And it isn’t a formality you sort out in an afternoon. On returning to Spain after five tax years abroad I went onto the special regime that applies to people who come back, and registering as self-employed meant giving it up. The sums did themselves and they were unpleasant: to be able to charge a few thousand euros I had to hand back a tax advantage worth considerably more.
I took the work on anyway, before knowing whether I’d be able to charge for it. And I haven’t charged: more than a year later I haven’t seen a cent from either of them. I haven’t pushed very hard either, because I had no way to issue the invoice or anything to back it with. When I do register, I’ll go and claim what I’m owed.
The third is that I don’t know how to sell, and on top of that I don’t like it. I know how to build the tool and I know how to explain it. What I don’t know is how to pick up the phone to call a stranger, or how to put a price on my own work without shrinking. On 20 May, discussing a proposal with an agency I was going to work with, I wrote: «The price of 10500 is out of proportion», and compared it with a proposal of my own that came to around 3,600. Someone put it this way in a forum and there’s no improving on it: «Almost everyone comes in way too low when they’re starting out».
In my case it wasn’t just charging too little. It was not knowing what what I did was worth, because I hadn’t yet done the whole of it even once.
And there’s one piece of evidence that admits no argument: the only thing I’ve ever delivered in my life was handed to me by a friend. Not one job has come through a door I opened myself.
With those three things on top of me, the strange part isn’t that I didn’t build a company. The strange part is what I did do.
The forty-seven minutes of 2 August
This is the part I don’t like telling, which is why it goes in whole.
The idea didn’t come out of nowhere. Around then I was coordinating a job for a client in Dubai whose development was being done by two freelancers I’d hired on one of those sites where you pay by the job. I talked to the client, they wrote the code. On 14 July I put it in writing like this: «the development has now finished and in the end it was executed by a freelance». It worked.
So the question that night was, in its own head, reasonable. On 2 August 2025, at 22:20, I described a business model: set up an automation agency, find clients close to home, and deliver the projects with freelancers hired over the internet. I’d be the one talking to the client and to the freelancers. And I added this:
In principle the client won’t know I’m working like this. Do you see the idea as viable?
Forty-seven minutes later I wrote the opposite:
And what if I say openly that I execute the projects with freelancers and I’m something like a consultant? In that case I would take on responsibility for the project
And eight minutes after that: «I’m starting to like this consultant/adviser approach, someone who understands the client and their need».
Nobody changed my mind. There was no conversation with anyone and no rule stopping me: I changed it myself, by writing it, on seeing the sentence set down in black and white.
I take two things from that night. The first is that the temptation existed, and anyone who says it never occurred to them is either lying or hasn’t tried. The second is more useful: writing things down is what makes me see them. Everything I’ve built since rests on that.
There’s a third detail that’s only visible from here. That occurred to me when I still didn’t know how to build. Today I wouldn’t do it for anything, and not out of scruples: I’d do it myself, because I can. The temptation to hide who does the work goes out on its own the moment you know how to do it.
Fifteen at once
On 15 August, at 23:45, I wrote the sentence that best sums up the mirage:
I want to launch 12 or 15 at once
They were small products for freelancers and small businesses. Six minutes later I was asking how much it would cost to have the development done outside.
Fifteen products at once are zero products. I know that now and I think I knew it then.
That was my monkeywork of 2025, and this time I set it on myself: manufacturing the props of a company —names, logos, avatars, decks, plans, indicators— instead of finishing one thing and getting paid for it. When there’s nobody waiting, the work stretches until it fills the whole night.
What was actually happening
All of the above is true and doesn’t change. But if I stop there, I tell the year wrong.
In 2025 I wrote to the machine, by hand and almost always at night, the equivalent of four books. That’s the figure that really tells the year: I didn’t build a company, but I didn’t stop for a single month learning to talk to these machines.
And underneath something changed without my noticing. In March I was asking it for documents —decks, letters, a plan. In November I was asking it to make something work on a server. That’s why halfway through the year I moved from ChatGPT to Claude, and started paying for the subscription: it wasn’t fashion, it was that what I was asking for had stopped being text and had started being software.
Along the way I paid for subscriptions to more than half a dozen platforms I don’t use today: website builders (Elementor, Webflow, Weglot), booking managers for a holiday-lets setup (Tokeet, Beds24, Amenitiz), flow builders (n8n, Flowise), a couple of things for making applications without writing code (Lovable, Suna), and scattered along the way Airtable, FeedHive and Scribe. I don’t call it money thrown away, even though some of it was. Each one taught me where the ceiling is on tools that promise to do it for you, which is exactly what I needed to know before setting about building for real.
And underneath, quietly, three things were happening that didn’t seem important at the time. On 17 June I asked it to analyse the videos of Casey Meehan, from the Blazing Zebra channel, who talked about working with AI in a way I hadn’t seen before. On 31 July, my questions were no longer about decks: they were about how to hand the context of one conversation on to the next. And on 8 September I passed the machine the link to a working session with these words:
Sharing the link to today’s Claude session so you can see what for me has been a productive and efficient session
That’s someone who has found a way of working and wants to repeat it.
Sixty-seven lines
There’s a part of the year that isn’t in any ChatGPT or Claude conversation, because it didn’t happen in a chat. It happened inside a code editor, and until two days ago I had it down as lost.
I started with the Claude extension inside VS Code and carried on with Cursor. Along the way I paid for the Blazing Zebra course —Casey Meehan’s channel, which I’d started watching in June—, where at last somebody explained techniques rather than tricks, and for research I used two others: Perplexity and NotebookLM.
None of the four taught me to program. They taught me to commission.
I learned it by doing it, at night: one in every three messages I wrote in there falls between ten at night and seven in the morning.
That’s the real price of the year and it doesn’t appear on any balance sheet.
And there’s one number that sums it up better than anything I could tell you: the editor keeps count of the lines of code it suggests to you.
Over twenty-one working days between November and December, autocomplete suggested me sixty-seven lines. The agent suggested me a hundred and ninety-six thousand six hundred and forty-two, and I accepted a hundred and thirteen thousand six hundred and thirty-two.
I didn’t write that code. I didn’t write any of it. Not one line, not even pressing a key to finish another.
What I did do was say no. Of every ten lines it suggested, I rejected four. That’s the only thing separating what I was doing from hitting accept until the screen stops complaining, and it’s also what saved me: the judgement to know what isn’t good enough wasn’t given to me by any tool, it was given to me by the previous twenty years looking at spreadsheets that lied.
The day I stopped asking for code
At the end of November something happened that I didn’t see at the time. My messages stopped asking for things and started handing out work.
I want you to prepare the prompts to give each agent once module A is finished. Remember to assign each agent a specialisation
you are the project manager of the project. Analyse the current state and propose next implementation steps to your team of developer agents
And two days after that I sent two separate jobs out at once, one for each half of the work, each with its specialist. Meanwhile I had a second agent reviewing what the first one wrote, and when the two didn’t agree, the one who arbitrated was me:
the reviewer agent has found an error and on suggesting a fix the agent asks me this. Analyse the answers and suggest a reply
That isn’t asking a machine for things. It’s a small crew, with a foreman and an inspector, and me as site manager. I was doing it in November 2025 and I didn’t put a name to it until much later.
I also learned to ask for options instead of answers. When I didn’t like a screen I didn’t argue: I asked for three different versions and chose.
I wrote it like this: I want you to show me at least three more variants, and above all guide me on how to compare them. Comparing was the work; choosing was mine.
And the other half has to be said too, which is the one I still had to learn: all the plumbing was still mine.
I moved the folders, I started the server by hand, I restarted it when it hung, I pasted the passwords into the chat so the thing could save the work. The crew built on its own; the one fetching the tools was me.
It took me another four months to realise that that was monkeywork too.
What I learned
That I can build the apparatus of a company in an afternoon, and that that isn’t a company.
That the only thing that went ahead all year was the one that had somebody waiting on the other end, and that everything else —the fifteen products, the agency, the beautiful deck— were comfortable ways of not reaching that moment.
That the part I’m good at, building, is also where I hide from the part I’m not. And that willpower doesn’t fix that when on top of it you get paid at the end of the month, can’t invoice, and feel awkward putting a price on your work.
That you can close a whole year of work with zero euros coming in and it still not be a wasted year. But you have to say both halves of that sentence, not just the one that sounds good.
That in parallel, in the real job, I was putting hours into the company’s flow tool and ended up connecting with the internal team that automates processes, despite working six hundred kilometres away from them. Nobody asked me to. And that was the year I knew that my professional path, inside or outside that place, was going to go hand in hand with this.
But above all, and this took me a year to see: 2025 wasn’t the year I didn’t build a company. It was the year I learned to work with these machines. Not by watching tutorials: by doing it. The decks that sold nothing, the plans nobody executed, the eight ideas of the spring and the platforms I paid for and abandoned were the syllabus. Expensive, disorganised and with no teacher, but a syllabus.
In the previous chapter the mistake was not sharing the tool. Here it was the opposite and more expensive: sharing the promise of the tool before having it.
What got me out of there wasn’t a better idea. It was a small job from a Canary Islands public body, with a date and with someone waiting on the other end of the email. That’s where I really learned, and that’s what the next chapter is about.