Do you still prompt ChatGPT task by task, or do you use one master document that the agent follows for the whole project?

I've been experimenting with a different way of working on larger projects. Instead of explaining the project again in every prompt, I keep a master document with the important rules, architecture, decisions, constraints, things that must not change, and the overall goal. Then when I give the agent a task, the idea is basically: Read the master document first Inspect the current state of the project Do only the requested task Don't contradict previously documented decisions Update documentation when something genuinely changes It feels much more consistent than having a huge prompt every single time, especially once a project gets complicated. But I'm also wondering if this creates its own problems. The document can become huge, outdated instructions can survive for too long, and the agent might spend too much context just understanding the rules before doing any actual work. For people using agents on real projects: Do you prefer a master spec/document like this, smaller task-specific prompts, or some hybrid between the two? And at what point does a "master prompt" become too big to be useful?