I used to think the problem was my prompts.
Every time an LLM lost track of what I was building mid-session, I'd rewrite the prompt. Add more detail. Be more specific about the architecture. Paste more code. The model would snap back on track — for a while — and then drift again.
After months of this, I finally diagnosed the real problem: I was prompting when I should have been looping.
The difference nobody talks about
Prompting is a one-shot act. You write a request, get a response, and start fresh next time. The model has no memory of what came before unless you manually re-paste it. Every new conversation is a blank slate.
