AI-assisted development, with every change reviewed

The first things I shipped with AI were built in Lovable: a few dashboards, then a concept for SAPinsider.org v2, then the Mastering SAP Collaborate event site. The Event Manager came next, also in Lovable.

Autonoma started there too, most of the front end at least, and then I moved into Claude, running it inside VS Code like an IDE. That’s where I’ve worked since.

Right now the thing on my desk is Lux, a macOS app in Swift. Think Alfred from Batman: one assistant, one conversation thread, and anything you want made or done starts there.

I don’t have years of Swift behind me. I have a clear picture of what it should do, and that turns out to be the part that matters.

What each tool is for

Lovable is for getting a front end on screen fast, while the idea is still moving. It’s how Collaborate got its first real shape.

Claude is for the build. I write the plan with it, queue up the changes, read what comes back, and go again.

Codex is the second pair of hands. When a diagnosis is finished in one tool, I have it write a handoff doc so the other can pick up cold, with no chat history needed.

I come from design. I know HTML and CSS well enough, and I don’t have decades of backend work behind me.

A year ago that gap set the ceiling on what I could ship alone. Now the ceiling is how clearly I can say what I want.

Where I stay in the loop

Two rules that don’t move. An AI doesn’t get root, and it doesn’t get unrestricted access to sensitive data.

Make the machine it runs on safe first. Then, inside that box, give it wide permissions. Tight permissions in a risky environment is the worst of both.

The review happens at the decision points, and there are more of them than people expect. Two from Lux:

The voice layer had started filling silence with small talk while the model behind it was slow. The obvious move was to roll the branch back a week.

We didn’t, because both failures were older than the branch. Three targeted reversions instead, plus one new rule: when the model is slow, Lux says “my brain is slow right now” and waits.

A scroll loop pinned the CPU at 100%. The first job was a diagnosis, with no code changed.

The first theory was wrong in two places, a second adversarial pass caught it, and only then did anything get touched. Then a handoff doc, so Codex could take it from there.

Where it breaks

Vibe coding lets a vision get built with far less miscommunication than any handoff to a team I’ve been part of. That’s the honest upside, and it’s big.

It breaks whenever clarity suffers.

A thin brief or a fuzzy definition of done, and you get a confident build of the wrong thing. When I look back at my bad weeks with these tools, the request was under-specified every time.

So the job changed. A year ago the slow part was turning feedback into a revision. Now that’s fast, and so is going from an idea to something you can see.

Most of my day is building in conversation, with an AI or with the people who need the thing, and letting the AI settle the details. Where it’s safe, I’m handing it more of the process, so I have better information in front of me when a decision comes up.

Real engineering is engineering that works, however it was made.

Collaborate has been live since May, writing each visitor a personal event guide in about 12 seconds. That’s the review.

---

Thanks for reading. If you’re working on something like this, email me.