plan-for-goal
Plan for Goal
Craft one prompt to hand to a coding agent's /goal command. The prompt is re-injected on every turn of the orchestration loop — it must hold up cold, without any "as discussed earlier".
The Topic
$ARGUMENTS
How /goal works
/goal is a self-perpetuating loop, not a one-shot task. The objective the user pastes in is re-injected on every continuation turn until the agent decides the work is complete. Each turn the agent re-reads the prompt from scratch, inspects the current state of the repo, and either keeps working or marks the goal complete after auditing every requirement against current evidence. The harness already enforces verification rigor, scope preservation, and budget discipline — your prompt's job is to set direction, taste, and a bar that's actually auditable.
A good goal is bigger than one prompt but smaller than an open-ended backlog — one objective with one stopping condition, sized so the loop can run for hours without your input but knows when to stop.
What the prompt must contain
Use the conversation so far as your input. The prompt you produce needs five things:
- A directional outcome. What's true about the world when this is done. Outcome, not a task list. Not "implement X", but "users can do Y and it feels Z".