Tasks

A task is one piece of work the project intends to do. Tasks move from to-do, through in-progress, to done, and can be archived when they stop being relevant.

Tasks are a record of intent. Nothing happens because a task exists — work happens when the agent is asked to do it. The value is that the queue survives the conversation: what you decided last week is still there this week, in the order you decided it.

Where tasks come from

• You write them. The backlog of things you know the project needs.
• The agent splits them out. A request too large for one turn becomes a set of steps, so each one can be verified on its own rather than reviewed as one enormous diff.
• Reviews and gates produce them. Findings that are worth fixing but not worth blocking on become tasks instead of being forgotten.

The point estimate

A task written with a description is sized when it is created: the same quick pass that gives it a title also judges how much AI work it looks like, and the task carries that as a rough point number — a simple question, a focused change to one component, or a page with its back end. Work that is bigger than a single slice is labelled a big task rather than given a number, which is the honest answer and usually a hint to split it up.

The number is approximate and assumes the standard model tier; hovering it shows the typical range behind it. It is there to be read before you open the task, so the queue tells you what it will take as well as what it is.

Ordering and due dates

The order of the queue is yours to set, and it is the main signal about what matters next. Due dates are optional and advisory — nothing is escalated automatically when one passes; an overdue date is information for you, not a trigger.

Tasks and turns

Working through tasks one at a time is also the cheaper path: one outcome per turn, reviewed and accepted, costs less than a sprawling request that has to be partly undone. See what a turn costs.

Next: Feature briefs · Your first Ask