Prompts

What the model is actually told, and how it is assembled.

Assembled per turn, not stored

There is no single prompt file. Each turn builds its own from parts: who the assistant is, what it has been told to remember about you, what was recalled for this question, what tools exist, and what has happened since you last spoke.

That is why the assembly stage in the agent loop has three boxes in front of it. By the time the model sees anything, the turn has already been priced, drained, and given its memory.

Manifest by default

The prompt lists what is available rather than inlining all of it. The assistant reads what it needs when it needs it, instead of carrying the whole library into every exchange. The result is a smaller, faster, cheaper turn that can still reach everything.

The workspace is part of it

Several parts of the prompt come from files the assistant maintains about itself — its identity, its standing instructions, what it has learned about how you work. Those are ordinary Markdown files you can open and edit. See Workspace.

Which tools it can see

Roughly half of what a turn costs before you have typed anything is tool descriptions. Every tool the assistant has is always callable, but only some of them are described to it on every turn; the rest it finds by searching, which costs one extra step and nothing else.

Which ones ride is yours to set. Conscience shows every tool, how often you have actually called it, and whether it is being carried anyway — that intersection is the one worth acting on, because a tool you never call and load every turn is pure cost. Turning one off there reaches your next conversation, not the one in front of you: see What a turn sends for why the list is held still while a conversation runs.

The same list is a file, config/working_set.yaml, if you would rather edit it in one pass. The names under core: are yours; the groupings and the description beside each name are written from the tools themselves, so they cannot drift from what the tools actually do. Everything not on the list is written underneath it, commented out, so adding one is uncommenting a line.

This changes what the assistant can see, never what it is allowed to do. Taking a tool off the list makes it slower to reach and no less permitted; permissions are a separate question with a separate answer.

One name cannot be removed: the search tool itself. It is how the assistant reaches everything not on the list, and without it, it can only reach what is.