The best input is a short product brief: who the product serves, which problem it solves, how the main workflow behaves, and what must be true for the build to count as complete. Describe the product boundary before assigning individual coding tasks.
Include five elements:
- User and job: Name the primary user and the outcome they need.
- Core workflow: Trace the main path from entry to success, including failure states.
- Data and integrations: List records, systems, APIs, and ownership rules.
- Constraints: State the required stack, security boundaries, and areas the agent must not change.
- Acceptance criteria: Give observable behaviors and executable checks that define done.
A product brief is not the same as a task prompt. The brief explains the system and its boundaries; a task prompt assigns one change within that system. Ask the agent to return assumptions, architecture choices, open questions, and a proposed task sequence before it edits code. Correct a misunderstood business rule at the planning stage, when the fix is still cheap.
Use one concrete user journey and one example input/output pair to remove ambiguity. If another developer could not sketch the same product after reading the brief, the agent probably cannot either. Start with one narrow workflow, approve the plan, and expand only after the first slice passes its acceptance checks.
Related reading: What makes a good AI coding-agent prompt? and What is Plan Mode in an AI coding agent?.
