The starting state.
A solo founder turned up at the studio one afternoon with a USB stick, a name (Thorim), and a working piece of software. The software was real. One hundred and six skills, thirteen specialist agents, fourteen commands. It ran, it worked, the founder used it themselves every day. The product was, in the literal sense, finished.
But the product was naked. No brand. No website. No payment flow. No licence system, no admin to track who had paid, no piracy controls, no affiliate workflow. The hard part was already done, and the easy parts were missing.
I had a month to turn that into a launched, paying product. That part is the story below.
What "AI-augmented" actually means in a small studio.
It is easy to misread the AI-augmented studio thing as AI did the work and the human watched. The reality is different, and worth saying out loud, because the misreading is what makes solo studios get into trouble.
AI is a force multiplier on the plumbing. It does not, today, replace taste, judgement, or the ability to say no.
The split that worked for me on the Thorim build was roughly this.
What I trusted AI to do end-to-end:
- Generate the first draft of every page of code.
- Build the licence-management backend (Stripe webhook, Cloudflare Worker, database schema for users, licences and installs).
- Write the boilerplate for the admin dashboard.
- Draft transactional emails, terms and conditions, the FAQ copy.
- Refactor, lint and tidy as the codebase grew.
What I had to do myself, every time:
- Pick the name and the metaphor. The founder brought Thorim; I shaped what it stood for.
- Decide on the visual direction. The cube logo did not come from a model. It came from rejecting eleven other directions first.
- Write the headline, the subtitle, and any line a real human will read carefully.
- Decide which features to ship and which to drop.
- QA every flow as a real user, on a real phone, with a real card.
- Make the calls about pricing, refund policy, and how aggressive to be with anti-piracy.
The pattern is simple. AI is fast at the things where direction is already clear. It is dangerous at the things where direction is still being chosen. So you use it after you have made the call, not instead of making it.
The thirty day timeline, week by week.
Week one. Brand and direction.
The first week was mostly not at a computer. Sketching, sleeping on names, looking at twenty AI tool websites and writing down what every single one of them did wrong.
The mistake every other AI tool was making was looking the same. Same gradient. Same hero. Same blue. Same vibe. The brief became: do not look like an AI tool. Look like a craft tool.
By Friday: a name backstory (a Norse creature of discipline and fortitude), a cube logo, a colour system (plasma purple as anti-blue, void black as anti-gradient), a voice rules card, a one-page brand guide. The brand was ninety percent locked. The rest was just applying it.
Week two. The marketing site.
A one-page launch site, but built around teaching instead of selling. The hero is a live head-to-head where the same prompt runs against a vanilla coding assistant and against Thorim, and the difference shows itself. No marketing copy could have done that work for me.
The page also has a browsable directory of every skill and agent, a 3D rotating cube that reveals the team on hover, and an admin-configurable pricing card and FAQ so the founder can update both without redeploying.
This is where AI moved fastest. Every component was a first-draft-from-prompt, then a thirty-minute pass to make it actually correct, then ship.
Week three. The licence platform.
The bit nobody wants to write themselves. Stripe checkout to issue a licence. A Cloudflare Worker to validate licences from the desktop app. Tables for users, licences, installs, refunds, gifts. An admin dashboard with sales totals, traffic charts, affiliate workflows, and a piracy alert when a single licence checks in from too many IP addresses in too short a window.
This was the most AI-led part of the build. The schema, the routes, the controllers, the migrations were all drafted by a model and tightened by hand. The control was in the code review, not in the writing. I read every line before it went live. That is the part no model does for you.
Week four. Polish, payments, launch.
Real money tests. Refund flow tests. Affiliate payout tests. Email deliverability tests. Photographing the founder's actual setup for one piece of marketing photography that ended up being the only image we needed.
Friday afternoon: launch tweet. First sale within fifteen minutes.
What 1,000+ purchases in a month actually proves.
It does not prove the brand worked, the site worked, or the platform worked. It proves the founder had already built a product people wanted, and could not, until that month, give us money for. The job of the studio was to remove that single blocker.
Not less, but not more.
What this means if you have a finished product and no surface.
If you are sitting on a USB stick of working software, or a recipe a chef has spent ten years on, or a service nobody can find online, the bottleneck is almost never the work. It is the surfaces. The brand. The website. The way a customer pays you and the way you confirm the order.
Those surfaces are where the AI-augmented studio can compress months into weeks, if the human in charge has taste, and a finished product to work with, and the discipline to say no to the model when it suggests something that is fast but wrong.
The order matters. Finished product first. Brand and surfaces second. AI on top of both. Skip step one and the speed becomes the problem.
A shorter version, for anyone scanning.
- AI does not replace taste. It accelerates execution.
- Use it for code drafts, boilerplate, schemas, lint. Do not use it for naming, direction, or strategic calls.
- A finished product plus a one-month AI-augmented sprint can ship a launch. A half-finished product plus the same sprint cannot. The model will not catch the gap.
- The work that disappears is the plumbing. The work that stays is the judgement.
That is what this kind of studio actually is. Not a place where AI builds the company. A place where a human who has decided what to build can ship it in a quarter of the time it used to take.