tl;dr Codex fixed homesareexpensive.com again, then cloned a Scrabble site in a few hours.
I had an idea at work for an at-home project – could you take Claude Code, point it at a website, and ask it to build a clone? Obviously the backend stuff would be hidden, and actually sharing the end result (let alone monetizing it) would be a legal and ethical violation…but as a personal project, it would probably lead to some interesting ideas. How would Claude Code work out functionality the website has, like search? Would Claude Code actually investigate the website’s JavaScript and keep its work as close to the original as possible, or would it try to recreate it?
It has been a little while since I gave up on Claude, which could be its own blog post…in short I am not opposed to giving it another go, but I was tired of subreddits like ClaudeExplorers and their tendency to anthropomorphize generative AI. I swithed from Claude Chat to free ChatGPT, then yesterday tried upgrading to ChatGPT Plus (costs $20 a month, same as Claude Pro) to get upgraded chat and Codex access.
It Fixed homesareexpensive.com
This part was not worth its own blog post because the fix was already shared here, but that time I was using free ChatGPT. This time I got to use the OpenAI agent, and I thought I’d ask it to help me with some scp commands. I wanted to have Smack’s full configuration, complete with certificate and Dockerfiles, in a folder so I could compare it to what is actually shared for us on GitHub.
It was taking forever, mostly because I lacked the foresight to use a series of commands to use scp but ignore things like frontend config files. Then I thought….wait a minute….I can just ask Codex to SSH in, have a look around, and authorize it to use similar commands as last time to renew the certificates.
Once in, it fixed the problem in a matter of seconds.
It Cloned Scrabble
Besides needing some extra passes for code formatting, what impressed me here is that when I asked Codex to look at a Scrabble site it “one-shotted it.” Yes, it ran out of tokens and I had to have it finish in the morning (making what I wrote somewhat misleading), but it still built something playable without a lot of guidance or pushback. What it created may even look and act differently enough from the source to be able to share.
The AI, it explained, simply pulls from a Scrabble dictionary and heuristics to make logical moves. The question is, is the AI smart enough to use AI?
Closing Thoughts
Throwing up something fully AI-generated is not impressive, except maybe to people who have been hiding from the news and social media for the last four years. The value here, I THINK, is in the ability to use the AI agent to then explain its work and diagram out architecture.
Or….not. This could go one of two ways. Either everyone promoting AI is right, and people using AI will defeat everyone in single combat….or the ones avoiding AI are right, and soon they will re-emerge gloriously to clean up the mess created by all these agents.
Having it fully AI-generate something could have been my first mistake. Typically there is a human to at least break things into logical chunks, or at least that is the cadence that I have found works. If you ask someone to generate something important, something with a lot of complexity, then it makes no sense to have them do a lot of it at once without proper review. Better to review it, line by line, and build it out slowly.
Or…not. Maybe even that could prove problematic in the long term. Maybe the only way to truly