They’re not the task for today. They’re the standing rules the AI carries into every task in the project, different from a longer prompt: a prompt is usually one request, project instructions shape every request inside the project.
Think about hiring a new employee for one department. The job description doesn’t say “answer this one email.” It says what the role is, what standards matter, how to communicate, what to avoid, and when to get approval. Project instructions are exactly that: “this is the job you’re hired for inside this project.”
How it shows up
In Claude Projects, ChatGPT projects, or a repo-based agent setup, they might say: write in our voice, never send outbound messages without approval, use plain language, and ask questions when source material is missing. Those rules don’t belong in every message; repeat them by hand and you’ll forget one, so they live in the project as the default posture. They work next to project knowledge: the instructions are the job description, the knowledge is the reference binder. They also sit near a system prompt, the deeper layer set by the product, while project instructions are the layer you control.
Why you care
The best project instructions are short enough to stay alive. A junk drawer of rules just adds noise, and a few strong ones beat two pages of vague preferences. When an instruction becomes a repeatable way to do work, it may belong in a skill instead. Good instructions save you from starting from zero every time and turn a blank chat into a trained seat at the table.