Building with AI: A PUBLISHING AGENT
Every Sunday night, a Slack bot named Larissa told me what to build. She'd done her homework over the weekend, looked at the search data, and come back with three options and a short argument for each. Monday morning I'd read them, and reply with a single word.
“/build”
And then she’d get to work. New calculators written, illustrated, published, with the documentation and the roadmap updated to match. I hadn't done any production work. I'd approved a suggestion from a piece of software that had decided, on its own, what was worth doing.
That inversion is the whole project. For most of my career the direction of work ran one way: I decided what to make, then made it. Here the agent did the deciding. I just had to say yes. (The single word is mostly there so I stay in control of what gets spent, on tokens and on time.)
The experiment was to make the research, the prioritisation and the production all happen without me, and that my entire contribution was judgment.
This is the full build-flow - from outside the machine room. I post in Slack during the week when I stumble on something that either was annoying (“I don’t get the US Cup system in recipes”) or something that looks like a good idea. Trying to mimick a reality, Larissa of course has a roadmap, she pairs the requests and braindumps from the week gone by with simple trend research, and a look in the Google Search Console data for the site.
It's just ugly calculators
Unitwiz is a site full of small calculators and charts. TV sizes, cooking conversions, how big a 55-inch screen really is. I started it because I was trying to work out exactly that, looked at the big calculator directories that rank for these queries, and noticed they make real money while being genuinely ugly. The thought was the obvious one: if these can win, surely I can build something nicer and take a slice.
The calculators were never the interesting part. The interesting part was the agent that made them.
The pages themselves are deliberately plain, close to unstyled. That isn't laziness, it’s just me leaning on what the space looks like. Someone arrives from a search with one question, wants the number, and leaves. The tool is the design, and restraint was a design decision.
How the agent earns its keep
The engine is a GitHub Action on a weekly schedule, a Slack bot, and the Claude API, wired to my Google Search Console data and Google Trends.
- Every Sunday it pulls the numbers, finds the gaps between what people search for and what the site already covers, and ranks the opportunities.
- Then Larissa posts the shortlist. On "/build," Claude generates the calculator and its supporting content against a strict design system I'd defined up front: page structure, writing rules, SVG patterns, SEO conventions.
- The page ships to Cloudflare, the docs update, the roadmap ticks over.
Larissa researches, reasons (admittedly in the simplest way), proposes, and builds.
20.000+
Impressions
0%
Average CTR (ok, there were 5 clicks)
The machine worked. The market didn't.
Over about ten weeks the site collected roughly 20,000 impressions. Google saw the pages, indexed them, and put them in front of people 20,000 times. Of those, 5 turned into clicks. Of a thousand ranked queries, 999 sent me precisely no one.
The automation was flawless. The pipeline never broke. The pages were clean, fast, and correct. But I had built a perfect factory in a town with no road to it.
The calculator strongholds I’m up against aren't impressive by looks, but they've been accumulating authority for fifteen years, and every search anyone runs reinforces it. You don't out-build that. You'd have to out-last it, and I’m not about to spend fifteen years ranking for "6 foot in cm."
There's a bigger possibility I can't rule out, though, and it's less comfortable than just picking a hard market. Sites like these are really just enormous Q&A pages: a question goes in, a number comes out. That is precisely the kind of query Google now answers directly, in the AI summary at the top of the results, without anyone needing to click anything at all. I may have built a very efficient agent for a way of finding things that the web is in the middle of leaving behind.
It’s alright it didn’t work. At least it made me believe in humanity again.
I could have left the machine running. Set it to "/build" on a loop and let it populate the internet with thousands of calculators nobody would ever find. But two things stopped me. First, what’s the point? Second, that would just be handing money to Anthropic. I like them, but not that much.
The deeper thing is that I'm probably a bit relieved it's this hard. I went in assuming I could glance at someone's revenue model and clone a cut of it without breaking a sweat. I couldn't. And a world where you can, where any working product can be trivially copied the moment it's spotted, is a worse and messier world than this one. Nothing would be safe and nothing would be worth building.
I want to believe good products earn their place over time, and that being easy to make was never what made them good. I can't prove it. Making is under real pressure right now, and I'm not going to pretend I know how that resolves. But I'm still in love with the craft of building things, and will keep digging into what “craft” actually means right now.
The point of the experiment was never to replace the making. It was to find out how much further one person could take it.
The job moved upstream
The part I'll keep isn't the calculators or even the failure. It's where the work moved.
I didn't draw the pages. I wrote the rules the pages had to obey, and then I judged what deserved to exist. That's the same job I've done with design systems for years, pushed one step further upstream. When production is nearly free, making things stops being the scarce skill. Deciding what should be made is the whole job.
Deciding what should be made is the whole job. And the systems for deciding are changing with every release.
Building with AI: A PUBLISHING AGENT
A weekend robot that built calculators while I watched. I wired up a Slack bot, the Claude API and a schedule. It researched, wrote and shipped pages on its own. An experiment in automating the making, and in how little it turned out to matter.
Every Sunday night, a Slack bot named Larissa told me what to build. She'd done her homework over the weekend, looked at the search data, and come back with three options and a short argument for each. Monday morning I'd read them, and reply with a single word.
“/build”
And then she’d get to work. New calculators written, illustrated, published, with the documentation and the roadmap updated to match. I hadn't done any production work. I'd approved a suggestion from a piece of software that had decided, on its own, what was worth doing.
That inversion is the whole project. For most of my career the direction of work ran one way: I decided what to make, then made it. Here the agent did the deciding. I just had to say yes. (The single word is mostly there so I stay in control of what gets spent, on tokens and on time.)
The experiment was to make the research, the prioritisation and the production all happen without me, and that my entire contribution was judgment.
This is the full build-flow - from outside the machine room. I post in Slack during the week when I stumble on something that either was annoying (“I don’t get the US Cup system in recipes”) or something that looks like a good idea. Trying to mimick a reality, Larissa of course has a roadmap, she pairs the requests and braindumps from the week gone by with simple trend research, and a look in the Google Search Console data for the site.
It's just ugly calculators
Unitwiz is a site full of small calculators and charts. TV sizes, cooking conversions, how big a 55-inch screen really is. I started it because I was trying to work out exactly that, looked at the big calculator directories that rank for these queries, and noticed they make real money while being genuinely ugly. The thought was the obvious one: if these can win, surely I can build something nicer and take a slice.
The calculators were never the interesting part. The interesting part was the agent that made them.
The pages themselves are deliberately plain, close to unstyled. That isn't laziness, it’s just me leaning on what the space looks like. Someone arrives from a search with one question, wants the number, and leaves. The tool is the design, and restraint was a design decision.
How the agent earns its keep
The engine is a GitHub Action on a weekly schedule, a Slack bot, and the Claude API, wired to my Google Search Console data and Google Trends.
- Every Sunday it pulls the numbers, finds the gaps between what people search for and what the site already covers, and ranks the opportunities.
- Then Larissa posts the shortlist. On "/build," Claude generates the calculator and its supporting content against a strict design system I'd defined up front: page structure, writing rules, SVG patterns, SEO conventions.
- The page ships to Cloudflare, the docs update, the roadmap ticks over.
Larissa researches, reasons (admittedly in the simplest way), proposes, and builds.
20.000+
Impressions
0%
Average CTR (ok, there were 5 clicks)
The machine worked. The market didn't.
Over about ten weeks the site collected roughly 20,000 impressions. Google saw the pages, indexed them, and put them in front of people 20,000 times. Of those, 5 turned into clicks. Of a thousand ranked queries, 999 sent me precisely no one.
The automation was flawless. The pipeline never broke. The pages were clean, fast, and correct. But I had built a perfect factory in a town with no road to it.
The calculator strongholds I’m up against aren't impressive by looks, but they've been accumulating authority for fifteen years, and every search anyone runs reinforces it. You don't out-build that. You'd have to out-last it, and I’m not about to spend fifteen years ranking for "6 foot in cm."
There's a bigger possibility I can't rule out, though, and it's less comfortable than just picking a hard market. Sites like these are really just enormous Q&A pages: a question goes in, a number comes out. That is precisely the kind of query Google now answers directly, in the AI summary at the top of the results, without anyone needing to click anything at all. I may have built a very efficient agent for a way of finding things that the web is in the middle of leaving behind.
It’s alright it didn’t work. At least it made me believe in humanity again.
I could have left the machine running. Set it to "/build" on a loop and let it populate the internet with thousands of calculators nobody would ever find. But two things stopped me. First, what’s the point? Second, that would just be handing money to Anthropic. I like them, but not that much.
The deeper thing is that I'm probably a bit relieved it's this hard. I went in assuming I could glance at someone's revenue model and clone a cut of it without breaking a sweat. I couldn't. And a world where you can, where any working product can be trivially copied the moment it's spotted, is a worse and messier world than this one. Nothing would be safe and nothing would be worth building.
I want to believe good products earn their place over time, and that being easy to make was never what made them good. I can't prove it. Making is under real pressure right now, and I'm not going to pretend I know how that resolves. But I'm still in love with the craft of building things, and will keep digging into what “craft” actually means right now.
The point of the experiment was never to replace the making. It was to find out how much further one person could take it.
The job moved upstream
The part I'll keep isn't the calculators or even the failure. It's where the work moved.
I didn't draw the pages. I wrote the rules the pages had to obey, and then I judged what deserved to exist. That's the same job I've done with design systems for years, pushed one step further upstream. When production is nearly free, making things stops being the scarce skill. Deciding what should be made is the whole job.
Deciding what should be made is the whole job. And the systems for deciding are changing with every release.
Building with AI: A PUBLISHING AGENT
A weekend robot that built calculators while I watched. I wired up a Slack bot, the Claude API and a schedule. It researched, wrote and shipped pages on its own. An experiment in automating the making, and in how little it turned out to matter.
Every Sunday night, a Slack bot named Larissa told me what to build. She'd done her homework over the weekend, looked at the search data, and come back with three options and a short argument for each. Monday morning I'd read them, and reply with a single word.
“/build”
And then she’d get to work. New calculators written, illustrated, published, with the documentation and the roadmap updated to match. I hadn't done any production work. I'd approved a suggestion from a piece of software that had decided, on its own, what was worth doing.
That inversion is the whole project. For most of my career the direction of work ran one way: I decided what to make, then made it. Here the agent did the deciding. I just had to say yes. (The single word is mostly there so I stay in control of what gets spent, on tokens and on time.)
The experiment was to make the research, the prioritisation and the production all happen without me, and that my entire contribution was judgment.
This is the full build-flow - from outside the machine room. I post in Slack during the week when I stumble on something that either was annoying (“I don’t get the US Cup system in recipes”) or something that looks like a good idea. Trying to mimick a reality, Larissa of course has a roadmap, she pairs the requests and braindumps from the week gone by with simple trend research, and a look in the Google Search Console data for the site.
It's just ugly calculators
Unitwiz is a site full of small calculators and charts. TV sizes, cooking conversions, how big a 55-inch screen really is. I started it because I was trying to work out exactly that, looked at the big calculator directories that rank for these queries, and noticed they make real money while being genuinely ugly. The thought was the obvious one: if these can win, surely I can build something nicer and take a slice.
The calculators were never the interesting part. The interesting part was the agent that made them.
The pages themselves are deliberately plain, close to unstyled. That isn't laziness, it’s just me leaning on what the space looks like. Someone arrives from a search with one question, wants the number, and leaves. The tool is the design, and restraint was a design decision.
How the agent earns its keep
The engine is a GitHub Action on a weekly schedule, a Slack bot, and the Claude API, wired to my Google Search Console data and Google Trends.
- Every Sunday it pulls the numbers, finds the gaps between what people search for and what the site already covers, and ranks the opportunities.
- Then Larissa posts the shortlist. On "/build," Claude generates the calculator and its supporting content against a strict design system I'd defined up front: page structure, writing rules, SVG patterns, SEO conventions.
- The page ships to Cloudflare, the docs update, the roadmap ticks over.
Larissa researches, reasons (admittedly in the simplest way), proposes, and builds.
The machine worked. The market didn't.
Over about ten weeks the site collected roughly 20,000 impressions. Google saw the pages, indexed them, and put them in front of people 20,000 times. Of those, 5 turned into clicks. Of a thousand ranked queries, 999 sent me precisely no one.
The automation was flawless. The pipeline never broke. The pages were clean, fast, and correct. But I had built a perfect factory in a town with no road to it.
The calculator strongholds I’m up against aren't impressive by looks, but they've been accumulating authority for fifteen years, and every search anyone runs reinforces it. You don't out-build that. You'd have to out-last it, and I’m not about to spend fifteen years ranking for "6 foot in cm."
There's a bigger possibility I can't rule out, though, and it's less comfortable than just picking a hard market. Sites like these are really just enormous Q&A pages: a question goes in, a number comes out. That is precisely the kind of query Google now answers directly, in the AI summary at the top of the results, without anyone needing to click anything at all. I may have built a very efficient agent for a way of finding things that the web is in the middle of leaving behind.
Metrics
20.000+
Impressions
0%
Average CTR (ok, there were 5 clicks)
It’s alright it didn’t work. At least it made me believe in humanity again.
I could have left the machine running. Set it to "/build" on a loop and let it populate the internet with thousands of calculators nobody would ever find. But two things stopped me. First, what’s the point? Second, that would just be handing money to Anthropic. I like them, but not that much.
The deeper thing is that I'm probably a bit relieved it's this hard. I went in assuming I could glance at someone's revenue model and clone a cut of it without breaking a sweat. I couldn't. And a world where you can, where any working product can be trivially copied the moment it's spotted, is a worse and messier world than this one. Nothing would be safe and nothing would be worth building.
I want to believe good products earn their place over time, and that being easy to make was never what made them good. I can't prove it. Making is under real pressure right now, and I'm not pretending to know how that resolves. But I'm still in love with the craft of making things, and that's exactly why I went this far into the tools, not the other way around.
The job moved upstream
The part I'll keep isn't the calculators or even the failure. It's where the work moved.
I didn't draw all the pages. I wrote the rules the pages had to obey, and then I let the agent run. That's the same job I've done for decades - amplifying the designers choices empowering the rest of the organisation to work faster. When production is nearly free, making things stops being the scarce skill.
Deciding what should be made is the whole job. And the systems for deciding are changing with every release.