Building Is Free Now. Asking Isn't.
Tuesday, 13 October. Six emails to clients, sitting in my Drafts folder. Finished. Each one a single click from sent.
I started thirteen build sessions instead. The emails stayed where they were.
That week I made 1,740 commits across my projects — the most of any week on my record. The week before, I'd made 1,536, the record until then. In the first half of that earlier week, seventeen emails to clients went out. Then, on the Wednesday, I wrote down four priced proposals to send by the Friday.
None of the four went. Two had reasons I could point to. For the other two, the only reason was me.
Before I go further, I should say what a commit means here. More than nine in ten of mine since late July list Claude as a co-author. I decide what gets built, I read what comes back, and I send it back when it's wrong — but most of the code isn't typed by me. That's not a footnote. It's what this essay is about.
Building Got Cheap
Building used to cost time I could feel. A day on a feature was a day not spent on anything else, and sooner or later that trade made me ask whether the feature was worth the day.
That trade has mostly gone. I pay a flat monthly fee for the AI that writes most of my code, so another feature costs me nothing I'll ever see on a statement. A session runs while I'm doing something else. Building has become the cheapest thing I do — and the one that's always there.
Free isn't quite the word. The AI still needs directing. What it produces still needs reading, and that takes real attention. But the cost that used to make me stop and choose — the cost of one more feature — is now so close to nothing that I don't notice it.
And building feels like progress. That's not the same thing.
In July 2025 the research group METR published a randomised trial of early-2025 AI tools: 16 experienced open-source developers, 246 real tasks, their own repositories. With AI, they took 19% longer. Afterwards, they believed the AI had made them 20% faster.
The tools have improved since, and METR now thinks developers are probably faster with them. But when it tried to measure that, it said in February 2026 that the new data was unreliable — mainly because more and more developers chose not to take part. They didn't want to work without AI. I can't say I blame them. I wouldn't have signed up either.
What stays with me from the first study is simpler. The people doing the work couldn't feel which way it went.
The Excuse That Didn't Survive
The easy story is that building and selling fight over the same hours. If that were true, the weeks I built most would be the weeks I sold least.
My record says otherwise. In the week of 14 September I made 997 commits — a record at the time — and twelve sends went to clients and prospects. The week after: 1,441 commits, and two new paying clients.
So the building wasn't taking the selling's time. When building costs close to nothing, it doesn't crowd anything out. It just sits next to whatever else I decide to do.
Which means that in October, when the priced proposals didn't go, I can't blame the commits.
I chose the build.
Why the Build Wins
In 2013, Paul Graham wrote that founders resist recruiting users one at a time because they'd "rather sit at home writing code than go out and talk to a bunch of strangers and probably be rejected by most of them." Back then, that preference had a price. The code took the founder's own hours, and eventually the hours ran out.
Now the code mostly writes itself while I watch. Staying home and building costs almost nothing. Sending a price to someone who might say no costs exactly what it always did.
I think that's the whole mechanism. In April I wrote that my gap wasn't a knowledge problem — it was exposure. To act is to be seen acting. Jeffrey Pfeffer and Robert Sutton, who wrote the book on the knowing-doing gap, traced it in 1999 to "the willingness to let talk substitute for action." In my case, the substitute is building.
A build shows me its finish line the same day. Whether it works or not, I'm the only one who finds out. An ask goes to another person. The answer takes days. It might be no.
A build can't turn me down.
I've seen this before, in The Anonymous Account. In 2023 I posted 251 tweets in four months under a name that wasn't mine — and eight replies under my own. The content was never the problem. My name on it was. A priced proposal is the plainest version of that: my name next to a number, sent to someone who's free to say no.
What Has Worked
I've tried the obvious fixes. Written triggers. If-then plans. Rules with a date on them. The ones that depended only on me mostly failed the same way the emails did. Nobody else was holding the date.
Two things have worked.
The first is a person waiting. By Thursday 15 October the six emails had been in Drafts for about a week, and I didn't send a single email that day. The one thing I did send, by WhatsApp, went to someone who'd first got in touch that morning. We spoke at lunchtime. I said I'd send some questions over, and I sent them that afternoon. Someone was waiting on something I'd promised that day.
Nobody was waiting on the six.
The second is a pile-up. On 29 September, five asks I'd been putting off went out in a single day — because the cost of everything they were blocking had finally grown bigger than the cost of the conversations. It worked. But it's an expensive trigger: one of those five had waited 24 days.
Both work for the same reason. Not asking finally costs something I can feel. A person waiting makes it cost early; a pile-up makes it cost late.
So here's the change. I stop being the only one who knows when an ask is due. I book the call before the proposal is finished, so the proposal has a deadline someone else can see. I tell people when to expect the price — and then the date is theirs as well as mine.
Counting the Other Side
My own dashboard, the one that gives me points for different kinds of work, had building beating the ask six to one last quarter.
That ratio flatters building twice. I did build more. But the dashboard also pays me for commits automatically, straight from GitHub, while an ask only scores if I log it at the end of the day. I built that dashboard myself. Without meaning to, I made it watch the work I already do — and wait to be told about the rest.
At Auto Alpha Advisory I've written about why most AI projects stall before they ship. That's the client's version: a pilot that never reaches production. Mine runs the other way. My projects ship. Then they stall, before I've asked anyone to pay for them.
Cheap building is good news. The thing I'm good at no longer costs me much. But it also means the volume doesn't prove anything any more. A week of 1,740 commits tells me the AI was busy.
The numbers that would tell me something are how many people I asked, and how many said yes. Those are the ones I've counted least.
I haven't fixed this. Publishing this essay is one of the asks.