Every action Frontier takes — a chat message, a background Agent run, a commit summary — sends a prompt built from a template. Those templates ship with sensible defaults and are yours to edit.
Templates use tokens that get replaced with real content at send time: the workspace's own prompt, its attached Domains, the thread's history, and so on. The editor lists every token available in each template, so you can rearrange or trim what gets included rather than guessing at the format.
This is the right place to change how much context each run receives, or to enforce a house style across every conversation at once. It is also the easiest place to quietly break things — a template missing the token that injects your workspace's context will still run, just with an agent that knows far less than you expect. Every template can be reset to its default individually, so experimenting is safe.
Two small things make editing them bearable. Any prompt field expands to fill the screen, so a long template is read as a whole rather than through a five-line window. And the message composer shows roughly how large the next run will be — the workspace prompt and its context, the conversation so far, and what you have just typed, added up — which is usually the moment you notice a thread has grown expensive long before the message you are about to send has.
If you only want to change context for one project rather than all of them, write a Domain instead — that's the per-project version of the same idea.