Skip to main content

​
Summary

This starter turns the customer-support case study into a small local workflow: read an inbound customer email, read a local support policy document, classify the case, and draft a safe reply for human review.

​
Status

starter

​
Why It Exists

Customer support is a practical local-agent shape because the useful context is often already nearby: an email export, a policy document, a refund rule, or an FAQ file. This starter keeps that boundary visible by requiring explicit local paths instead of pretending the agent knows the policy from memory.

​
Folder Structure

customer-support-email-agent-starter/
├── index.mdx
├── SOURCE_NOTES.mdx
├── skill/
│   └── SKILL.md
└── src/
    ├── email_triage.py
    ├── policy_loader.py
    └── reply_guardrails.py

​
Quick Start

This is a starter, not a finished helpdesk integration. The code sketch uses only the Python standard library and focuses on the local-path boundary. Example: 
from reply_guardrails import draft_policy_grounded_reply

result = draft_policy_grounded_reply(
    email_text="Subject: Refund request\nI received the wrong item.",
    policy_path="/Users/example/support/refund-policy.md",
)

print(result.reply_subject)
print(result.reply_body)
print(result.needs_human_review)
For a repo-level smoke check, run python3 scripts/verify_example_projects.py from the repository root.

​
Included Sample Files

  • skill/SKILL.md: skill instructions for checking customer email and drafting policy-grounded replies from a local document path
  • src/email_triage.py: lightweight classification and summary helpers
  • src/policy_loader.py: local policy loading and evidence extraction helpers
  • src/reply_guardrails.py: draft generation and human-review guardrails

​
Constraints

  • No mailbox, Gmail, CRM, or helpdesk adapter is implemented.
  • The starter reads local text-like policy files only.
  • Drafts are intended for human review, not automatic sending.
  • Classification is keyword-based and intentionally small.

​
Next Steps

  • Add a mailbox adapter that writes inbound messages to explicit local paths.
  • Add structured policy sections with stronger retrieval.
  • Add evaluation fixtures for refunds, billing, complaints, and escalation.
  • Add an audit artifact that records policy evidence and reviewer decision.