I have been using ChatGPT pretty heavily for coding and building projects, and something that keeps tripping me up is context management.
At the start, everything feels smooth. I explain what I am building, set some patterns, maybe even define structure, and the outputs are solid. But after a while, things start to drift:
It forgets earlier decisions
It rewrites parts of the code I did not want touched
It breaks consistency in naming, structure, and patterns
It sometimes confidently moves in a direction that does not match the original intent
I usually end up either trying to remind it of everything again or cleaning things up manually afterward.
I have tried a few approaches to fix this. I have kept structured markdown files with context, decisions, and architecture. I have also experimented with tools like Speckit and Traycer to define specs and guide outputs. Breaking tasks into smaller prompts helps a bit too, but it slows things down.
They help, but it still feels like I am fighting context drift more than I should be. At some point it becomes less about building and more about constantly re aligning the model. I am curious how others are dealing with this in practice. Are you maintaining some kind of external source of truth, resetting context frequently, or sticking to one long thread. Also, are there any workflows that actually scale beyond small tasks.
Would love to hear what is working or not working for you.
[link] [comments]