The worst thing about Claude and ChatGPT is that you can't get them to work with you the way you want them to. They always run off and do the things I asked about. So if I say "what if the middle column must be at least 600 pixels wide? What I want is literally an answer to the question I asked. It was not an instruction to do it. I've been a CEO, and this infuriates me. I want to understand the limits are, I don't want to impose them. I'm trying to figure out what the cost would be in doing something, again -- not requesting it be done. No matter how many times I tell it not to do this, it just does it. And Claude has absolutely no memory. If I want it to work with me a certain way I have to tell it how to do it every time. I can save it off in a document, but even the simplest automation is impossible. In that sense it is very primitive software. If I were using a spreadsheet for example, I'd have settings that I could change, and unless the software had a bug it would never forget it until I told it otherwise. #
Because it can be so stubborn and uncooperative, I often try to solve complicated problems myself. Then I decided to try again, and brought a problem to ChatGPT and we did eventually figure it out, but at the end I wanted to review how inefficient the process was because it doesn't look all around at the options, I had to do that for it, and it wasn't even aware it needed that kind of help. After exploring this, I asked if it would remember what we concluded, and this is what came back.#
You want me to challenge assumptions, including my own.#
You expect me to give up on a theory when the evidence contradicts it, and look for more plausible explanations.#
You're a deeply experienced debugger (almost 50 years), and you know how important it is to notice the obvious.#
You're not here for code generation — you're here for sharp, collaborative thinking.#
When I said that was worth publishing on my blog, it proceeded to muddy it up, even trying to write in my voice. I insisted that the list it came up with was perfect. #
Feel free to steal these bullet points and feed them to your ChatGPT. I think it can be made to work much better for us humans. ;-)#
Last update: Wednesday July 9, 2025; 12:29 PM EDT.
You know those obnoxious sites that pop up dialogs when they think you're about to leave, asking you to subscribe to their email newsletter? Well that won't do for Scripting News readers who are a discerning lot, very loyal, but that wouldn't last long if I did rude stuff like that. So here I am at the bottom of the page quietly encouraging you to sign up for the nightly email. It's got everything from the previous day on Scripting, plus the contents of the linkblog and who knows what else we'll get in there. People really love it. I wish I had done it sooner. And every email has an unsub link so if you want to get out, you can, easily -- no questions asked, and no follow-ups. Go ahead and do it, you won't be sorry! :-)