Run tasks

Confirmation cards

Skills no longer ask clarifying questions — a turn starts as soon as you send it. The one card that can still pause a turn is the approval card for hard-to-undo actions.

Updated

No clarifying questions

Every skill starts working from your message as sent. Anything the request leaves open — which databases, what date range, which venue tier, the participants or the primary measure of a design, the grouping variable of an analysis — is settled by the agent from the request and from what the literature typically does. It states the assumptions it made in the first line of its chat summary (and, for a design, in the design-basis section), so you can redirect it in the next turn.

To steer the result, put the settings in the request itself: "RCTs only, 2015–2026, PubMed and OpenAlex", "office workers, N-back accuracy as the primary measure". The skill guides templates show the shape of a well-specified request.

The approval card

The one card that can still pause a turn is the approval card. It appears only before an action that is hard to undo — deleting a file, overwriting an uploaded original, excluding participants or rows from a dataset, or sending data to an external service — and shows the code or file involved.

  • Approve and proceed runs the action; the same kind of action is not asked again in this session.
  • Cancel stops that action; the agent continues without it and tells you so.

While the card is waiting, the composer placeholder says so. You can review or revoke the approvals granted in a session under Customize.

Cards in other languages

Card text follows the session language. In Korean, Japanese and Chinese sessions the options and the placeholder are translated.