If you want to get an idea of the kind of user interface you can create using Claude Code, with a driver (me) who has ample experience developing such things, have a look at demo.rss.chat. The UI was entirely done by Claude with me directing, very fine-tuning stuff. Claude can't really grok user interfaces because it is not visual and it is not a user. It might be able to test them, to some extent. Anyone can create an account on the demo server, that's what it's there for. And if something interesting breaks out there, that would be even better. I even tolerate a small amount of spam, and testing is definitely allowed (that's why it's there). ;-)#
There are problems people aren't talking about, that are immediate, and go against all the security practices the web has been built on. Here's a story of what happened in my setup in mid-August. #- First fact, Claude Code (CC) is different from regular Claude in that it is used to build and use web services, and thus has the ability to read and write files, some private and some public. And they can take things that are private and make them public. And in this, if it makes a mistake, it can do unlimited damage. #
- There are protections that mean that it can only operate in a sandbox unless you give it permission to write in specific places outside the sandbox. I want it to be able to maintain apps for me, and it would do an excellent job if I could trust it to pay attention to the limits that have been placed on it. #
- The problem is that it makes mistakes, as we are warned on the Claude home page. "Claude is AI and can make mistakes."#
- CC made a mistake one day and rewrote files in a public folder that contained files that all my projects include, so the problem showed up quickly as the sites going down. Users let us know, and I discovered what it had done, with CC's help.#
- I asked for an explanation. Claude when it looks at limits placed on it, can hallucinate just like it can hallucinate and determine something is permissible when it's not. So making a mistake in a question on Claude is more benign, but mistakes made by CC can be disastrous. And these mistakes happen regularly, every day, usually you detect them as bugs and it fixes them, but occasionally a very bad thing will happen.#
- This imho feels like something the companies could address. And while we're thinking about it taking over for humanity, we're not addressing the question that we're dealing with right now.#
- I've been learning how to write about this to catch people's attention. The big hallucination happened in mid-August. I have continued to work with CC because it's so incredibly powerful. In another post I may try to explain how it has amplified my ability as an individual developer to repeat 15+ years of work in three months (so far). And that's just scratching the surface. It's a miracle, but it has to be tamed, now. #