GPT 5.5 is the most reliable model I’ve used, but I am more inclined to want to keep working if the agent is able to communicate naturally, and preferably with its own little ”style” of talking layered in small bits of the instruction. It’s been easy to tweak with pretty much all models I’ve tried, except for these recent ChatGPT models. It has the most cemented, seemingly hard-coded dialogue habits I’ve seen and it makes communicating about the code near unbearable sometimes.
I’d say top 3 worst slop-habits are: One-liners paragraphs with sentence length variation. It can’t stop saying ”sludge”, ”pressure”, ”severe”. And worst of all is the pattern it can’t stop itself from if given enough space: a set of one-liners as a pre-amble to the topic where at least hals are ”it’s not X, it’s Y”, ending on a declarative ”that is Z.” -conclusion statement that never actually convey anything concrete.
I’ve managed to instruct other models well on language, but nothing actually takes effect for longer than a sentence here. If anyone has managed to land on s style that actually gets consistent, I’d love to see examples! I love working with the model otherwise.
FYI: on codex desktop it will barely change style in the slightest. Hooked up to other harnesses where it’s more of a blank slate, it only seems to latch onto repeating gimmicks, not language patterns.
Honestly? I just wanna see if ANY kind of more natural-speaking style can be stumbled into here, because I will not have my smartest AI agent yet lose their performance in spending half of their tokens on language reminders. It has to be easy for the model or the style will just be a harm, I think.