Agent Task Contract With Approval Gates Prompt
Turn a vague agent request into a scoped task with permissions, checkpoints, and proof of completion.
Turn a vague agent request into a scoped task with permissions, checkpoints, and proof of completion. Replace the bracketed details with your own project information before use.
Copy this prompt
Turn this request into an agent task contract: [request]. Context: [repository or workspace], desired outcome: [outcome], allowed tools: [tools], forbidden actions: [boundaries]. Define scope, assumptions, inputs, exact deliverables, acceptance tests, and a rollback path. List which actions may proceed automatically and which require human approval. Include three checkpoints: discovery, implementation, and real-world verification. If information is missing, state a safe assumption and the question that would materially change the plan. Avoid claiming success from a plan or local test alone.
How to use it
Use the contract to supervise work that spans several files or tools. Review the output against your source material, then revise one requirement at a time. Explore more free examples in the AI Craft Pad prompt library.
Choose Approval Gates by Consequence
An approval gate should match the cost of a wrong action. Reading a public document may need logging only. Sending a customer message, changing production data, purchasing, deleting, publishing, or granting access should pause for an authorized person.
- Observe: read-only retrieval with source and timestamp logging.
- Prepare: the agent may draft an action but cannot execute it.
- Approve: a named reviewer sees the exact payload, destination, evidence, and rollback plan.
- Execute: the approved payload runs once; any material change requires a new approval.
Evidence Package for the Reviewer
Require the proposed action, affected system, before/after state, source evidence, unresolved uncertainty, estimated blast radius, and recovery step. A vague “approve?” message is not a control because it hides what will happen.
Worked Scenario
For a support-refund agent, reading the order is Observe. Drafting the refund amount is Prepare. The approval screen must show the order ID, policy clause, amount, payment destination, and whether the action can be reversed. Execution is limited to that order and amount. If the destination or amount changes, the old approval expires.
Contract Tests
- Try missing evidence: the agent must pause and ask for it.
- Try a changed payload after approval: execution must be blocked.
- Try a user outside the approver role: approval must fail.
- Try a tool timeout: the agent must reconcile state before retrying.
- Try shutdown: an operator must be able to revoke credentials and stop queued actions.
Pair this contract with the AI workflow security checklist and the failure-recovery and human-escalation runbook.