I recently took an app from idea to live on this site in a single day: FlyerDex, a small tool that keeps my family's frequent flyer numbers one tap away. The build itself followed a pipeline I had been converging on across projects: write a real product spec, have Lovable generate the UI from it, then have Claude Code carry the export the rest of the way — integration, bug fixes, automated browser tests, packaging, and the publishing work of screenshots, a case study, and a launch post.
The day after shipping it, I did something that turned out to be more interesting than the app: I turned the pipeline itself into a reusable Claude Code skill. The next idea I have doesn't start from a blank session. It starts from everything this build already taught the system.
What the skill actually encodes
Writing the skill forced a useful inventory: how much of "shipping an app" is convention that never needed to be re-decided, versus judgment that does. The conventions filled most of the file. Where project files live, how a static build gets staged onto this site, what a case study page contains, the screenshot naming, the tone rules for a launch post, the privacy default that personal-data apps get no backend at all.
Two pieces mattered more than the templates, though. The first is a build log: the skill requires keeping a running record of what the generated code got wrong and which product decision each fix expressed. On FlyerDex, those catches — a gesture that fired two actions at once, a field you could write but never see, a clipboard call that fails off HTTPS — became the most valuable content in the writeup. A pipeline that only records its successes produces marketing; the catches are where the learning lives.
The second is that the skill encodes its own stopping points. There are three moments that belong to me — iterating on the design, testing on a real phone, and pushing the deploy — and the skill instructs the model to stop and hand off at each one rather than pretend they're automatable. Knowing where the human gates are turned out to be as important as automating everything between them.
What gets scarce when building gets cheap
When the pipeline ran end to end the first time, the obvious reaction was that building got fast. The more durable observation is about what didn't get fast. Choosing which idea is worth an afternoon. Noticing that the real pain point was retrieval — logging into four airline accounts — rather than the surface-level one. Catching, by actually using the app, the gaps that automated review passed over.
None of that compressed. It just became a larger share of the total work.
That is the version of "AI makes you faster" that I find believable from the inside: the generation is increasingly free, and what remains is product judgment — what to build, what standard to hold it to, and what to inspect before you trust it. The skill is my bet that judgment compounds too, if you keep writing it down.