One assistant. Your context. Your choice.

The most useful assistance starts with understanding what you are trying to do. It also needs to understand where its permission ends.

Kai's assistant is designed as a shared layer across the product family. Agent gives it a dedicated home, while the consumer apps provide focused places to ask for help. Kai is pre-launch; the examples here describe the intended experience as we develop it.

Begin with the request

A request to prepare for a race may involve a running schedule, strength work, nutrition, and recovery. A request to reflect on a difficult day may need only the words you decide to share.

More context is not automatically better. The relevant question is which information helps with this request and whether you have allowed that use. If goals conflict, the assistant should explain the tradeoff and help you choose a priority.

The same principle applies to ordinary tasks. Asking Kai to compare two plans should produce a comparison. Asking it to save a workout should result in a clear account of what was saved. Asking for an idea should leave you free to decide what happens next.

Memory should be understandable

Remembered preferences can make repeated conversations easier: the equipment you have, a dietary constraint, a goal you want to keep in view. Useful memory also needs controls.

Kai's direction includes the ability to inspect, correct, limit, and remove remembered information. Product boundaries matter here. A shared assistant does not imply a universal transcript or unrestricted access to every private history.

Cycle context used for recovery does not become a dating signal. A private reflection does not become public content. Reading a message does not authorize retaining it as memory or sending a reply.

Know what actually happened

Assistance should distinguish a suggestion from a saved change, a planned activity from a completed one, and a successful task from work that remains unfinished.

That distinction matters even more when a request has several steps. If two parts are complete and one is interrupted, the result should make that clear. Resuming a task should not duplicate a purchase, a message, or a record.

For communication, publication, and purchasing, permission needs to match the actual action. Preparing a cart and buying its contents are different outcomes. A changed price or unavailable item may call for a new decision.

Support that fits your rhythm

A single conversation can be enough. Other goals benefit from a review or reminder over time. Ongoing assistance should have a scope and frequency you choose, with clear ways to change it.

You should also be able to use ordinary product controls directly. Manual programming, food logging, and private reflection remain valuable without starting an AI conversation.

We want assistance to make the products more useful while keeping your choices visible. The person asking for help remains the person deciding what matters.