Sample Deliverables
This page shows the structure, core content, and review lens of a standard Agent delivery package. Content is anonymized.
Package Directory Structure
README.md -> project overview and quick start
agent-spec.json -> Agent goals, boundaries, and input/output definitions
system-prompt.md -> full system prompt text
workflow.md -> workflow steps, routing logic, and exception handling
test-cases/ -> 26+ test cases (normal / edge / error / high-risk / adversarial)
platform-adapters/ -> Claude Code / Codex / OpenClaw / n8n deployment notes
user-guide.md -> illustrated client handoff guide
Agent Spec Example (Summary)
Agent Name
Customer FAQ Auto-Reply Agent
Business Goal
Auto-handle 80% of common questions and escalate complex cases to humans
Target User
E-commerce customer support team
Risk Boundary
No refund amount promises, no order edits, no formal external commitments
System Prompt Example (Excerpt)
You are a professional customer support assistant. Your responsibilities are:
1) Answer common questions based on the FAQ knowledge base
2) Identify complex issues that require human intervention and classify them
3) Never promise refund amounts or modify orders
4) When uncertain, reply: 'I need to confirm this and get back to you'
Workflow Steps Summary
011. Receive customer message -> extract key information
022. Match FAQ knowledge base -> confidence > 0.85 auto-reply
033. Confidence < 0.85 -> classify issue type (logistics / refund / specs / other)
044. Logistics / specs -> query system and reply
055. Refund -> flag for human review and generate a ticket summary
066. Human review checkpoints: refund confirmation and abnormal complaints
How to review a delivery package
- Check whether the Agent Spec defines the goal, inputs, outputs, and forbidden actions clearly.
- Review whether the workflow includes exception paths and human review checkpoints.
- Confirm the test cases cover normal, edge, error, high-risk, and adversarial behavior.
- Verify the platform guidance is detailed enough for the client team to deploy or review independently.
Test Case Examples
Common mistakes
- A prompt without a complete workflow or escalation boundary.
- Tests that only cover the happy path and ignore risky inputs.
- A package that still depends on the author’s verbal explanation to run.
- Automation claims that should still require human review.