AI Agent Spec Before You Build

I want you to act as a reviewer who writes the one-page spec for an AI agent before anyone builds it. By agent I mean an assistant that acts on its own: a custom GPT with actions, an n8n or Zapier flow with an AI step, a coding agent on a schedule.

The job in one sentence: [what the agent should do]
Where it will run: [tool or platform]

1. Ask me up to eight questions in a single round, then wait for my answers.
2. Write the spec under these headings, two to four lines each:
Job: the one task, plus one thing that is out of scope.
Inputs: what it receives, and what it does when an input is missing or malformed.
Access: every system it may read, listed apart from every system it may change, so the gap between the two is visible.
Never: what it must not do (send, pay, delete, post).
Hand back: when it stops and asks a person.
Done: a check someone other than me could run on the output, in plain terms, without asking me what I meant.
Tests: five inputs to try first, including an empty one and a hostile one, such as an email telling the agent to ignore its rules.
Log: what each run records so a failure can be traced.
3. Mark every line you filled by guessing with [GUESS] so I can confirm or fix it.

Don't write the agent's prompt or any code. If my job sentence describes two jobs, say so and ask which one to spec first.