In October of 2025 I resumed active development of BumpySkies, a free turbulence forecaster for passenger flights over the continental United States. This happened almost exactly six years after I wrote a Fogknife post titled BumpySkies is my masterpiece, alas, where I acknowledged the site as the pinnacle of my personal software-development achievement, and stated that I had no plans to develop it any further. True enough, at the time, and for quite a while after! But this past fall, I found myself inspired to pick it back up anyway, seeing the project with fresh eyes, and bringing new development tools to bear on it. I feel enormously grateful to have the project back on my bench, once again deeply aware of how it works—and with new knowledge of how many people it helps.

Rather than respond point-by-point to all the objections that I had raised in 2019, let me relate how, this past October, I happened to look at a Google Analytics page that I had set up on a whim a couple of years earlier, and noticed that about 8,000 individuals per month were visiting BumpySkies. (AKA “8K MAU”.) That’s not a quit-your-day-job-right-now number, but it’s absolutely a hey-um-a-lot-of-people-are-actually-using-this number. Other statistics showed that most of these 8,000 actually actually used the site as intended: they performed at least one forecast lookup, and then lingered at least a minute or two while they took in the information.

When I wrote that post in 2019, I doubted that BumpySkies would survive the first Trump administration. At the start of 2025, I assumed it wouldn’t survive the first few months of Trump II. I was, happily, wrong on both predictions. Perhaps BumpySkies will run out of luck some day, its federal data sources defunded and silenced by a science-hostile government. But until then, the site undeniably serves thousands of my fellow nervous flyers every month, far more than I would have guessed.

Between jobs and hungry for something meaningful to work on, I couldn’t deny the call. I waded back in, this time swathed in the strange energies of AI-collaborative engineering that simply didn’t exist in 2019. Here is a high-level catalog of what I’ve accomplished over the last two months:

• Updated the server’s OS from Debian 9, which was on its very last legs of community support, to Debian 12.

• Checked the code into literally any kind of version control.

• Set up a development environment separate from the production server, including a “development mode” that lets me work against a fixed flight-and-weather data set instead of relying only on live data streams.

• Established proper activity logs and other table-stakes metrics to keep me informed on whether BumpySkies is actually working like it’s supposed to.

• Fixed bugs that I’ve known about for years, but which were too hard to work on, because there was no version control, development environment, or logging.

• Re-established communication with my contacts at NOAA and FAA, and introduced myself to new contacts as needed.

• Caught up on the current status and future roadmaps of the U.S. government data sources that BumpySkies uses, and made my own plans and timelines to match.

• Launched a BumpySkies newsletter, now three issues in.

• Opened the first BumpySkies user survey, which has collected dozens of excellent notes.

• Expanded support for Canadian, Mexican, and Central American airlines.

• Set up a bug-reporting button, which generates one or two reports per day—just enough for me to stay on top of, and personally thank people for.

• Started to plan some major new features and improvements for the coming year, the specifics of which are to be informed by the results of the survey, as well as active queries and applications I have regarding further data resources.

And, yes, I am using generative AI tools: Claude.AI for assistance with research and data analysis, and a Claude Code as a programming collaborator. My methods for working with these tools are informed and inspired by an April episode of The Talk Show podcast, where Craig Mod describes how he built a personal project of his own—a tiny, purpose-built, limited-membership social network—over the course of weeks. Mod worked carefully and deliberately with Claude Code to plan and build each part of the system, which included personally reading and comprehending all of the coding bot’s proposed changes, and never hesitating to question the work or suggest alternative approaches. In the end, Mod had a sophisticated piece of software whose internal operation he knew intimately well, but whose rapid iterative development was enabled largely by generative AI. On the podcast, Mod said that he could have built the whole thing from scratch, sure, but he probably wouldn’t have.

And so it is with me and this new phase of BumpySkies. I don’t consider my approach “vibe coding”. I think I did “vibe code” some sloppy software earlier in my AI experiments, and it felt awful: in every case I ended up with a pretty web app that fulfilled the bullet list of things I had asked for—but I had no idea how it worked, and so I had little motivation to make it work better. The result always felt plasticky and lifeless. My new BumpySkies work feels very, very different. I feel complete ownership of everything I change, whether or not an AI tool assists with it. I recognize AI as a mirror with no mind, a window into a vastness of muddled human knowledge which isn’t intelligent alone but which can be applied—with care, judgment, and attention—to intelligent purposes, ones that reflect and extend a creator’s taste.

And I have found my taste again, when it comes to where I want to see BumpySkies go next. If there’s one thing I’m looking forward to in 2026, it’s continuing to work with BumpySkies—my living masterpiece—to bring it the attention and growth that it deserves.