Agentic Customer Support Copilot — AI skill for Claude Code

Support Copilot

A live, interactive AI customer support agent — built with the [Claude API](https://docs.claude.com), with a working chat demo, not just a script.

Live demo

**[Try it here](https://claude.ai/artifact/SAcYXTdW1DBzW79r9zKYon)** — note: trying the interactive part requires a free Claude account to sign in with (this is a platform requirement for any page that calls Claude live, not something specific to this project).

The problem

Support teams handle a high volume of requests that are genuinely simple to resolve — a refund within policy, a password reset, a straightforward cancellation — but still take a person's time to read, verify, and answer. The goal isn't to replace judgment on hard cases; it's to handle the clear-cut ones confidently and instantly, and know when to hand off the ones that actually need a person.

What it does

Given a customer message, the agent:

1. Checks a simulated account and a short policy knowledge base — its only source of truth. It's explicitly instructed never to invent an account detail or policy that isn't in that data.
2. Shows its reasoning as it works — a visible step-by-step trace ("Checked account details," "Searched knowledge base for refund policy") appears before the reply, so the process isn't a black box.
3. Decides: resolve directly, or escalate to a human — policy-covered requests (refunds within the window, resets, cancellations) get resolved immediately; genuine anger, ambiguity, or anything the knowledge base doesn't clearly cover gets escalated instead. A badge on each reply shows which happened.
4. Remembers the conversation — every message includes the full conversation so far, not just the latest line, so follow-ups and confirmations ("yes, please go ahead") are understood in context instead of the agent losing track and escalating out of confusion.

A bug worth mentioning

Early versions of this had no conversation memory — each message was sent to the model in tot