Two versions of this story. This is the story one. If you'd rather have it as a step-by-step system you can run today, here's the tactical cut.
Okay, so here's a sentence that shouldn't be true. I'm a marketer who has never really shipped a line of code, and over the last few weeks I built a thing that takes an idea, tells you honestly whether it's garbage, and then, if it isn't, builds it and tests it on people for you. It's called Forge. It's open source and free and you can go take it apart at the bottom of this page.
But the tool isn't the interesting part. The interesting part is the two years I spent being spectacularly bad at this first, and the exact moment it flipped. So let me back way up.
Picture the most eager intern who has ever lived. Brilliant, tireless, read the entire internet, will build you literally anything you describe in about four seconds. Incredible hire. There's one problem with him, and it's a big one.
He is completely incapable of telling you your idea is bad.
You say "build me an app that does X" and his eyes light up and he goes "GREAT idea, love it, on it," and he never once stops to mention that X already exists for free, or that nobody on earth wants X, or that you described X wrong. He just builds. That intern is every AI coding tool right now, and for two years, he and I worked exactly the way you'd expect two people with no impulse control to work.
I called it prompt-and-pray. I'd open the AI, paste in a giant wall of text describing the thing I wanted, hit enter, and pray. Then I'd fix what came out. Then I'd fix the thing my fix broke. Then I'd fix that. Round and round, and it did eventually produce something, and about half the time the something was a beautiful, functional product that literally no human being wanted, including me. I'd built it blind. The intern never told me not to, because telling you not to is the one thing the intern cannot do.
My first product came straight out of that loop. Split-Test-O-Matic, STOM for short, a split-testing tool I built entirely by hand with ChatGPT, back when it was running GPT-3.5 and was, to put it kindly, really stupid. The MVP took me two days, which for something like STOM was genuinely fast. Perfecting it took over a year. A year of me and a dim little model, one prompt at a time, grinding out something that should not have taken that long. Hold that thought, I come back to it.
I used to think my problem was that I couldn't code. Turns out that was never the problem. My problem was that I was letting the eager intern run the whole company, and the eager intern has no judgment, only enthusiasm.
Here is the thing that took me embarrassingly long to see. There is a whole pile of work that happens before "build it," and the intern can do almost all of it, and almost nobody asks him to.
He can go read what real people actually say about a problem, in their own words, off real forums and reviews. He can build you a fake focus group out of that reading and run your idea past it. He can pick your angle and test your words. And, if you set him up right, if you basically hire a second, meaner intern whose whole job is to disagree, he can look at your beloved idea and say the four words that save you a month of your life: "Don't build this. Here's why."
Everybody has an AI that will build anything. Almost nobody has an AI that will tell them to stop.
I didn't design all of this on a whiteboard. I stumbled into it across three builds, and each one handed me one piece of the puzzle. So, the three builds.
The first was a dinky little tracker I built for a boring weekly task I had to keep doing. Unglamorous tool, real need. But before I built it, on a whim, I ran the idea past a fake audience first, a simulated room of the actual people this was for, just to see.
And it came back and told me to rework it. Not "amazing work Adam, ship it." Rework. And the annoying part is it was right.
That was the click. For the first time, a machine in my process had told me something I didn't want to hear, and it had done me an enormous favor by doing it. I sat there and thought: the eager intern will never do that on his own. So I have to build the meanness in on purpose. Every project after that got a step whose only job was to try to talk me out of it before I wasted a week.
The second piece came from Nudge, a day planner I built for my wife. She's blind, so it talks back to her and runs on voice, and it had to actually work, not demo-work-for-thirty-seconds-then-fall-over.
That build got real. Sixteen phases. Somewhere north of 260 tiny saved checkpoints, each one a real step I could undo if I broke something. And what Nudge beat into me was this: a build needs a memory outside my head. It has to write down where it is at every single step. Because I am doing this between everything else going on and two kids with one on the way, and I will close the laptop for three days, and when I come back I cannot afford ten minutes of "wait, where was I." The project has to remember for me.
Then came Deadreckon, and Deadreckon is the one where I looked up and realized I didn't have three lucky projects, I had a repeatable machine and had just been running it by hand every time.
Here's what happened, and I still think it's a little nuts. I took a photo of a page out of an Army land-navigation field manual. The actual paper one. I handed the machine that photo as the single source of truth and said build me a working offline navigation app from this. And then I ran it through the full gauntlet. A synthetic Army review board tore it apart first. Then five straight rounds of usability testing with simulated users, scored every round, dragging it from a four out of ten up to a nine. Then I walked outside and used it against a real map in the real world and wrote down the three little things that annoyed me, and fixed those.
I had a working, field-accurate navigation app by the end of a lunch break. From my phone. And it was genuinely good, not good-for-a-guy-who-can't-code good, because it got put through the same beating a real product gets.
One authoritative input. Run it through the gauntlet. Ship it that day.
That was the moment. Three builds, three pieces: be honest, keep a memory, run the gauntlet. I was doing the same dance every time and re-inventing it every time. So I wrote the dance down and turned it into a tool, and that tool is Forge.
Forge is that order of operations, packaged so tightly that I only have to show up for the judgment calls and the machine holds everything else. It runs in stages, and every stage has to earn its way into the next one or the whole thing stops.
Before it does any research at all, it makes me write down what would make me kill the idea. That order matters more than it looks. If you decide your bar after you've seen the evidence, you will bend the bar until it says yes, every time, because you want the yes. Writing the kill line first is me protecting myself from myself.
Then it validates. It reads what real people say, builds the fake focus group out of that, runs the survey, and then, my favorite part, it turns loose that second mean intern whose entire job is to kill the idea. Not to be encouraging. To find the reason this is a dumb plan. And at the end it hands down one of three verdicts: don't build this, build it just for yourself, or build it for a real market. And it is flat-out not allowed to say "build for market" just because the fake customers clapped. It has to show real evidence of real people spending real money, or it caps the verdict and tells me why. The enthusiasm of imaginary people is not a business.
If the idea clears that, it builds, using the same disciplined engine with the memory that Nudge taught me. It writes the words as part of the build, not bolted on at the end, because on most software the words are half the product. Then it tests the thing on simulated users over and over until it scores a nine, with a blind user in the room every single round, because if my wife can't use it, it is not done. Then I take it into the real world, because no simulated user can feel glare on a screen or gloves on their hands or a GPS quietly drifting. Then it ships, and it writes down what it learned so the next build starts smarter than this one did.
The don't-build-it intern is the first move almost nobody makes. Here's the second one, and it's the one that makes the stuff actually feel good instead of just functional.
After the research comes back, Forge goes back and re-reads what people actually said and pulls out the things they need but would never put on a feature request. On that little tracker, people were quietly anxious about privacy, so "your data never leaves your computer" became a real, built feature, not a marketing sentence. On Deadreckon, not one person asked for automatic pace counting or terrain hazard warnings, but a soldier out in the field genuinely needs them, so in they went.
That's the whole difference between a tool that does what you literally said and a tool that feels like it actually gets you. And you only ever find that second layer if you go and listen before you build, which, surprise, is the same listening that told you whether to build at all. The research pays for itself twice.
You do not need to become an engineer. I am living proof, and I'm not being humble, I truly cannot write the code. What you need is to know the order of operations, and to know how much the machine can carry for you before the building part even starts.
Read what real people say. Ask the fake customers. Let something honest try to kill the idea. Then build. Then hand it to real humans and watch. Do it in that order and the endless thrash just evaporates, because the thing that was causing the thrash, building blind with the world's most agreeable intern, is gone.
Here's the number that still gets me. STOM took me a year to get right, by hand, back when the model was dumb. With this order in place, I could rebuild the whole thing in an afternoon. Maybe less. Same product. A year of my life against one slow afternoon. Not because I got smarter. Because I stopped building blind.
I packaged the exact order I use into Forge and put the whole thing out in the open, so you can run it, break it, and steal from it. It's free. Go take it apart.
I send what I'm building, what worked, and what broke. Free, no fluff.
Free. Unsubscribe anytime.I'm a marketer who builds his own tools. More at adamgarceau.com.