I am banging my head against the wall. We as users are not being treated seriously. I know this forum is not a complaint form, but seriously, allow me to at least vent.
1: The models are constantly fed PREVIOUS prompts into their CURRENT prompt input. That’s not right. They keep thinking that I am asking for a change that is already done. Why is it happening? Why did the engineers decide this is better over the way it used to work? It’s absurd. Why are they being fed previous prompts as current prompts? I have to open up every conversation with every minute change.
It happens more often over long conversation flows, which could be because the conversation chain is distilled into an abbreviated version of it and fed back to the model in order to alleviate some of the context window limitations? I have no idea, but that’s the only thing I could come up with and I’d rather have it forget the beginning of the conversation if that’s meant to be the solution. I’m just spitballing here.
2: And the incessant tool overuse… they forget that they can directly modify code with their built-in functions. So they keep painfully editing files with cat and sed, wasting tons of tokens and time. That’s ridiculous.
3: Then, they will inject multiple responses into their chain of thought. They’ll respond, then do some other stuff, like run a linter, and then respond again. Which results in them not summarizing what they’ve accomplished, instead they will reply that “build is ok and linter returned no errors”. I suspect that they believe that their internal chains of thought could be user inputs.
Those are all recurring bugs that are related to the harness that the Build mode utilizes. Not the models themselves, not custom instructions, not my prompts. It’s about the all the unnecessary, misleading data the models receive besides my actual prompt.
If anyone has custom instructions that help mitigate those 3 issues, please share. I tried, believe me, I tried.