Write a spec an AI agent can actually follow

I want an agent that {{goal}}. It has access to: {{tools, APIs, data}}. It runs {{when and how often}}.

Write the agent spec:
1. OBJECTIVE — one sentence, with the success condition stated as something checkable.
2. INPUTS — what it receives, in what format, and what to do when a field is missing or malformed.
3. PROCEDURE — numbered, deterministic where possible. Mark each step as tool call, judgement, or output.
4. TOOL RULES — for each tool: when to use it, when not to, argument constraints, and the maximum number of calls.
5. STOPPING CONDITIONS — success, give-up, and escalate-to-human, each with a concrete trigger. Include a hard step limit.
6. OUTPUT CONTRACT — the exact structure, with an example, and what to emit on failure.
7. GUARDRAILS — what it must never do (send external messages, spend money, delete, act on unverified data) and what requires approval.
8. FAILURE MODES — the five things that will go wrong, and the specified behaviour for each.

Assume the model will do something unexpected. Where does that get caught?