Open a project and every new chat starts inside it: its instructions apply, and the assistant can search its knowledge collections. Switching projects never moves a chat you already have; the change applies from your next new chat. The project chip above the message box shows which one is active.

1. Hand it over. A name, what you want done, what you want back, which of your notes to read (optional), how often, an end date, and the most it may spend in a month.
2. It plans first, when you say so. When you click Plan it, a frontier model writes a plan and the platform prices it. Planning costs about one to three cents and counts against the objective's limit.
3. You approve. Nothing else runs until you approve the plan, its limit and its end date. Changing the goal, raising the limit or extending the end date asks again.
4. It works in the background on your schedule, whether or not anyone is signed in. Each run gathers material (the searches in your plan and the notes you named), writes the draft, sends the material and the draft to a checking model from a different maker, and puts the result in your Notes.
5. It writes the method down. After three runs of the same plan, with the last two passing their checks, the method becomes a skill note you can read and edit.
6. It replays the skill. Captured once on a frontier model, written down as a skill, replayed on a cheaper route, every replay checked by a scorer from a different maker, falls back to the frontier model when a check fails; if time or the limit does not allow that in the same run, the draft is delivered marked as not passing and the next run starts on the frontier model. Whether a replay saves money is measured run by run, and it can be a loss.
7. You see what it cost, run by run, with what the gateway measured and what is an estimate labelled as such.

Where your content goes. Your goal is sent to Anthropic to plan it. Each run sends the approved search queries to Tavily; sends the search results, the notes you named and the draft to the writing model (Anthropic, or xAI through OpenRouter if you chose Grok); and sends an excerpt of the material and the draft to OpenAI for the check.

What it does not do. It never sends messages, posts, submits forms, buys or books anything on your behalf; where your goal needs one of those, you get a ready draft. It does not sign in to websites or read anything behind a login, and in this release it does not read web pages directly or connect to your email, calendar or accounts. Nothing is trained: the skill is a written note. It is not available on a deployment locked down to the local model.