ChatGPT-5 is out: Is it brain-dead, or are we?

 

5 simple (but weird) ChatGPT-5 tricks to get a 10x better response

ChatGPT-5 dropped last week, and it seemed as if my environment switched to the new one in mid-session for me. Suddenly everything seemed to go in the wrong direction.

For example: I had been experimenting with GPT-4’s Agent Mode, and after everything switched to GPT-5, agent mode was just gone. As in, nowhere to be found. After I move everything from point-a to point-b the old fashioned way (copy-pasta), I move on from what I was working on… “later problem” as my kid likes to say.

Then I turn to an internet draft document I’ve been working on and ask it to generate appropriate footnotes for one of the sections – and it comes back with a citation from the very draft I’m working on, making the entire citation self-referential and circular.

It just seemed like suddenly, ChatGPT became a little brain-dead (either that, or I did):

“How stair?” pic.twitter.com/XIpa6kBgcC

— AxisOfEasy (@axisofeasy) August 16, 2025

My wife sent me 5 simple (but weird) ChatGPT-5 tricks to get a 10x better response via D-Squared (Dylan T Davis) on Youtube, and I’m glad she did, because it explained the fundamental shift in architecture that has occurred in the new version:

One of the main “things to know” about ChatGPT 5 is that all the different models have been largely  consolidated and GPT-5 uses a routing engine after the prompt to choose which model gets selected:

Further, you can be more precise about steering it toward the right model through your choice of words – and even adding structure to your prompts via XML tags (it’s easier than you may think).

I also learned about OpenAI’s prompt optimizer via this video – which I wasn’t aware of beforehand (in the past I’ve been using Claude to optimize and generate prompts for GPT and Lovable) – and I like Claude for coding, I’ll be testing this OpenAI optimizer for non-code based tasks.

It turns out, you can also use the LLM engine to create it’s own rubric and then iterate against that rubric until it gives you it’s optimized response at the end of it.

There’s a lot packed in to a 10-minute video and I went ahead and subscribed to his email course to see what’s up with that (I have several separate canary email domains to receive all these newsletters and opt-in magnates I fall for).

But if you were wondering if “something changed” with the roll-out of GPT-5, you’re not alone and this video will help you understand why.