As we’ve been building Aboard, I’ve been very focused on our way of doing software—which is organized around business and organizational needs, and setting up data-backed apps in a reliable way. Our offering works hard to fit in with the world of product and process. 

I love our approach (obviously), and we’re learning a ton from early customers, but sometimes you gotta cut loose. So I took a quiet Monday morning and decided to check in with Claude Code and hack a little. 

Here’s what I built—a single slider that lets you explore all past and future time:

It’s also on GitHub if you want the code.

Why “Time Tourist?”

No matter what the news is on a given day, I like to remind myself that the world was here for billions of years and will be here for billions of years. We’re just time tourists.

I also like to watch videos on YouTube. So I thought to myself: What if I put those two things—time and videos—together?

As I coded, I kept a log of how long things took.

1. A minute: I told Claude Code to keep things chill and use basic web technology, not a bunch of weird stuff.
2. An hour: We chatted about how time can be converted to numerals, and after an hour, I had a somewhat buggy slider. The slider went from 13.8 billion years ago (big bang) to 10^78 years in the future (heat death), but “zoomed in” around the middle.
3. Ten minutes: I had it add eras, eons, and major events from Wikipedia. Claude is definitely better than it used to be at this. It sets up little virtuous loops of self-improvement, and fixes its own bugs.
4. Twenty minutes, but failed: Then I asked it to go find a bunch of YouTube videos for each event—and it whiffed it. Claude just doesn’t have the juice when it comes to searching the web or organizing thoughts.
5. Two minutes: I figured Google would know something about YouTube videos, but when I asked Gemini to dig a bunch up, most of them didn’t work.
6. Ten minutes: I went over to ChatGPT o3 and got it to make me a list of YouTube videos to correlate to all the events and eras, and it did a great job. I brought that back to Claude.
7. Two hours: Claude started to wear out. We crossed the Prompt Productivity Threshold, or Context Exhaustion. Eventually I got into the Bad Place and it told me that the Black Death happened around the same time as the Heat Death of the Universe.
8. Two minutes: Claude ran out of money and told me I couldn’t have any more code for many days, so I gave it more money. Upgrading sure is fast and easy.
9. An hour: A bunch of dumb little tweaks and fixes, dealing with it being too smart and trying too hard.
10. 20 minutes: Getting it live and deployed to GitHub pages.
11. A few more hours, but without paying much attention: I tried to get ChatGPT to up its game and add more videos. It was very enthusiastic but continually said it had added thousands of lines to a spreadsheet but then added around 20. I was really disappointed.

This last part—making a better, more robust set of links and videos—just kept producing bad results. ChatGPT couldn’t generate the file. Then it told me it was having kernel problems. It kept asking me to upload a sample file, again and again. I finally gave up; I’ll try again later. It seems to be an interlocking set of failures—internal timeouts, problems searching the web, and so forth. I guess next I’ll try Perplexity.

Nonetheless, in the end I’d call this a success. It’s cool that it works at all, and I was able to hack around, and I like the product it produced. The major problem was as the code kept growing, I needed to break everything into atomic, AI-sized subproblems and take them one by one—which needs a lot of time and thought. When you think about the future of coding, maybe that becomes a part of the process: A point where the bot recognizes bad loops are happening and decomposes the entire project into a set of much smaller tasks, then prompts the human to take them one at a time.

It’s odd to think that success may come when the bot is prompting the human, not the other way around. I’ll let you decide what that means, or if I’m right or not.

In the end, it makes me happy that we’re doing what we’re doing at work. Making these tools behave, using AI to kick things off and then letting humans take it from there, working on simple, reproducible patterns for making software ship faster—that’s progress.