AI Customer Service

Problem

A common online-support failure is that a model answers company-specific facts such as pricing, inventory, refunds, or contracts from memory and creates unsafe commitments. This skill constrains customer-service behavior to a local llm-wiki knowledge base, fits an OpenClaw or Hermes runtime, and returns traceable support decisions through WeCom, web, or API channels.

How It Works

The skill reads SCHEMA.md, index.md, and log.md to understand the product domain, support rules, and recent changes. It then classifies the user intent, retrieves pages covering products, pricing, troubleshooting, orders, invoices, and refunds, and chooses a response path: answer, ask one clarifying question, redirect back to company scope, or hand off to a human. It produces a concise customer-facing reply plus an internal decision object, letting a channel adapter hide internal fields and record content gaps. It also enforces boundaries for credentials, payments, contracts, and security incidents: it does not request passwords, verification codes, or card numbers, and it does not promise discounts, shipment, or compensation.

Boundaries

This is not a WeCom webhook implementation and does not own signing, encryption, idempotency, or message delivery; a channel adapter should handle those concerns. It is also not intended for medical, legal, or investment advice. When the wiki is missing, pages conflict, or the question affects purchase, billing, legal, or safety decisions, the correct path is human handoff rather than free-form model answering.