This is a quick article, wanting to note that as the context increases, it definitely becomes harder to work with an ai coding assistant.
If you’re giving it simple instructions, or have simple pages, it responds very well. It is able to fill in a lot of blanks, and it usually works. But the TFH application is showing me that the longer I work in 1 page, and the bigger that one context becomes, it also increases the difficulty of correctly prompting it to do what you want.
I’m currently hitting 25% context left with GPT-5, which is a new low for me. It was hard for me to go below 60% on other pages, but here it happens more and more frequently the past few days.
A few minutes after I started writing this, I also hit the usage limit for the first time. I spent an entire morning (08:30 to 11:15) prompting, and haven’t been able to make a single good commit, since Codex kept breaking stuff and not doing what I wanted it to do.
Something that seems to help with the context getting low, is just restarting codex and introducing it to a fresh new context. It loses it’s own mind due to it’s assumptions and thoughts (aka it’s context).