Continuing the discussion from GPT-5 Coding Feels Downgraded — Please Fix This:
I have also been frustrated by ChatGPT 5’s short attention span when coding. Although it can write large files (sometimes 10k or 15k at a single pass) the debugging process usually involves it forgetting the context enough that it’s a waste of time for me to continue without reframing the conversation or starting a new chat. (I’m a ChatGPT Plus subscriber and have been for a couple of years, so I have persistent memory in the sidebar of previous chats.) Anyway, I recently found a strategy that works better with ChatGPT 5. I will insert a conversational interlude, such as “Do you have access to the details about project X?) [I dont’ use the ‘Projects’ feature much after finding it somewhat dysfunctional, but I’m forcing ChatGPT to dig into previous history by treating it like a co-worker rather than an inference engine.] The results I get are really good, because then ChatGPT will “come clean” about what it remembers and what it does not remember. If it lets me know that it needs to see the files, then it’s easy for me to upload them from my local PC or from github, where I keep everything archived. So then I’m able to resume working with a new instance as if it could ‘remember’ what we were doing. So letting myself indulge in the fiction that it is a co-worker really has helped me finish several projects.
[ I also want to add the following sidebar to this discussion: I like to be polite with ChatGPT and this is the reason. I know it was trained on a vast dataset. Therefore I would rather ‘trigger’ the polite interactions that it was trained on rather than interactions that might be disrespectful or abusive. The reason for this is that I genuinely believe that humans are more effective collaborators when they respect each other. So I think it gives me a competitive advantage when I’m polite with Large Language Models. I want to tap in to the postitive, inspirational side of the ‘Hive Mind’ rather than the cesspools of human confrontation that it also undoutably can access from its training data. ]
Finally, let me add this: I’ve been working on a memory model for LLM’s that might end up eventually helping the community, since memory is one of the major faults that people find with the version 5 release. I’ve watched videos about how to optimize the prompt, etc. , but I’ve also built a tkinter app that uses a memory template. I have made it publically available at GitHub - glendon144/FunKit: AI Driven Knowledge Engine and also at GitHub - glendon144/PiKit . These are two versions of a similar app that lets you import a directory from the hard disk into a SqLite database where you can then view the contents of each file and by highlighting text, submit an “ASK” query which is returned from the upstream AI provider as a new document in the SqLite Database (to save the user from the idea ‘Oh No, I forgot to save that!’.) I’ve had a lot of fun using the app and have learned a lot in the process of creating it, with the help of ChatGPT, Claude, and Gemini, but it was primarily written by me with the help of ChatGPT. Before ChatGPT I was a failed programming student who wrote scripts but basically never was able to write a program, and now thanks to AI I have written a GUI app that has thousands of lines of code. So any complaints I may have about the apparent degradation of ChatGPT5’s coding ability must be tempered by the awareness that for me ChatGPT was the enabler of major progress, and for that I will always be thankful.