Aiming High and Schleps

Claude writes all my code now. If software is so much easier, products should be higher quality and solve more difficult problems. I’ve noticed I’m having a hard time taking the implications seriously.

For example, I wanted to build a bot to help me search for apartments in East Village. In a prior world that would have required web scraping and figuring out how to avoid bot detection on StreetEasy. It would have felt like an accomplishment to even get something like this working. Now, with tools like Browser Use, you basically just point a vision model at the website and give it a prompt. The Browser Use Agent looks at the webpage + DOM/Accessibility Tree and extracts the relevant data. It also gives rankings based on my high level visual preferences (e.g., modern finish in the bathroom). I showed Claude the docs, described the problem I was trying to solve, and got something working in a few hours. There are still Captcha issues, but the project basically works if you’re willing to babysit a little. This would have been crazy 2-3 years ago.

The upshot of getting this working so quickly was that I realized I built the wrong thing. I don’t actually want the agent to look at the listings for me. It’s helpful, but it’s not the hard part. I want somebody else to handle all of the annoying text, email and scheduling stuff for me. Basically I just want to be able to look at/edit the inventory, show up for tours, and sign the forms I need to start a lease. I experienced this with a “personal stylist” bot I built too. Finding clothes was easy, but I still haven’t bought them because the annoying part is getting the right sizes, deciding on a budget, and filling out checkout forms.

The apartment experience I’m describing feels like a schlep because it involves other people and I don’t want to spam them with my AI or have something slip through the cracks. The stylist project means I need to trust the agent with my money.

But, taking a step back, solving these kinds of reliability challenges is what will make a product compelling going forward. Any engineer can build the thing I did in a day or two, so a slapped together version that mostly works for me is not nearly enough. If I want anyone else to use the things I make, I have to dig into the annoying parts. The Captcha stuff, reaching out reliably, not missing any emails or texts across channels, having intuitive interfaces for interacting with the agent (e.g., text, email, Claude Code?), and so on. All these kinds of things that would have been pushed now need to be solved up front.

It may be easier to write code, but it just means I have to aim higher and do more schleps!