One founder, no dev team.
Claude was the dev team.

The constraint: one person, no dev team, a real problem that needed solving fast

I'd just watched a rebrand chew through weeks of admin time that should never have needed a human at all. Headers, footers, tables and body text across hundreds of documents, all wearing the wrong logo. I knew exactly what the tool needed to do. What I didn't have was a development team, a runway that could carry one, or months to spend learning to build it from scratch myself.

That's the constraint most solo builders actually face. Not "should I use AI to build software," but "I have a real problem, a limited amount of time, and no one else to hand the coding to." Claude became the answer to that specific constraint, not a philosophy I'd decided on in advance.

That loop repeated for every feature, every fix, every refinement to how the tool handled a stubborn table style or a footer that lived on its own XML part. It was closer to running a very fast, very literal junior developer than to typing a wish into a box and getting a finished product back.

Where AI-assisted building earned its keep

Speed was the obvious win. Features that would have taken me days to hand-code came back working in hours, which meant I could test an idea against real documents the same day I had it instead of a week later once the excitement had faded.

Iteration was the less obvious win. Because generating a new version of something was cheap, I could try three approaches to a stubborn formatting problem instead of committing to the first one that looked plausible. And more than once, Claude caught something in my own spec I'd missed - a batch-size assumption that would have broken on a genuinely large document set, or a formatting rule that only held for the tidy templates I'd been testing with. Those catches came from being asked to actually implement the thing, not just talk about it.

The line I drew: AI helped build the tool, the tool itself isn't AI

This is the distinction I keep coming back to, because it gets confused constantly. The build process and the product are two different things. My build process leaned heavily on Claude. The product that resulted reads a document's actual structure and applies a brand to it using fixed rules, every time, the same way. No customer's content ever passes through a language model to get rebranded. That line was a deliberate engineering decision, and it's one I'd make again regardless of how the tool got built.