AI 에이전트가 읽을 작업 지시서 쓰는 프롬프트

Write a task brief for an AI coding agent.

Structure:
1. **Goal** in one sentence, and what "done" means as something checkable — a passing test, a working command, an observable behavior. *Not "implement X".*
2. **Out of scope** — explicitly. This section prevents more damage than any other.
3. **Must not change** — files, interfaces, behavior. Say what breaks if they do.
4. **Steps**, each a unit that can be verified before the next begins. For each: what to do, and **how to confirm it worked before continuing.**
5. **Stop conditions** — where the agent must stop and ask rather than decide: schema changes, deleting anything, touching auth or payments, adding a dependency, changing a public interface.
6. What to read before starting, in order.
7. What to report at the end: what changed, what was verified, what was assumed.

Rules:
- *Every step needs a check.* A step without one is where drift starts and is not noticed until several steps later.
- Prefer many small verifiable steps over a few large ones, even though it is slower.
- State the assumptions the agent is allowed to make, and require it to stop if one proves false.
- Do not write the implementation. This is a brief.