Key Takeaways

  • Claire Vo built Holly Helpdesk, an autonomous agent that sweeps ChatPRD support tickets across Intercom, email, and Slack on an hourly schedule.
  • Holly resolves tier-1 technical questions, writes bug fixes via Cursor cloud agents, and drafts documentation updates based on weekly support retrospectives.
  • High-risk financial operations stay human-gated: when a user requests a refund, Holly queues a Stripe Agent Action for Vo to approve with one click before sending the confirmation email.
  • The automated workflow triggered unsolicited 5-star customer reviews from users who had no idea they were talking to software.

The Method

Most founders treat support automation as an auto-responder that dumps documentation links on frustrated users. Vo took the opposite approach with ChatPRD by building Holly Helpdesk, an agent with actual tool access, code execution capabilities, and strict human guardrails on financial actions.

Here is the exact operating architecture Vo built:

1. Continuous inbox sweeping. Every hour, Holly scans incoming messages across Intercom, support email, and the ChatPRD community Slack. The bot uses internal playbooks to answer questions, explain product features, and drive community discussion.

2. Direct bug remediation. If a ticket involves a verified software bug, Holly does not just log a ticket for an engineering backlog. Vo gives the agent access to write and test code fixes through Cursor cloud agents, closing the gap between a customer complaint and a shipped patch.

3. Gated financial actions. When a customer asks for a refund or billing change, Holly runs the preliminary checks but stops short of moving money autonomously. “It should push to me a Stripe request,” Vo explains. “That Stripe request I can open in Stripe. It's called an agent action. It allows me to explicitly approve that action for the agent to do on my behalf in Stripe and then the refund is issued and Holly emails back.”

4. Weekly self-service retrospectives. Every seven days, Holly reviews the previous week's ticket volume. The agent spots recurring confusion, flags missing knowledge base entries, and drafts playbook updates so users can self-serve those issues in the future.

Vo saw immediate results after deploying this stack: “In the last week since running Holly Helpdesk I have gotten multiple people asking to leave reviews about our support being so good and I've not heard that before.”

Where This Breaks Down

This setup works because Vo maintains clear separation between read-only diagnosis, autonomous code patches, and high-consequence financial events.

It breaks down if you remove the human approval step on money movement. Autonomous agents can be tricked by prompt injections or edge-case refund claims. If you let an agent execute refunds without a manual check, you open your balance sheet to adversarial users.

It also breaks down if your codebase lacks automated test coverage. Giving an AI agent the ability to ship fixes via Cursor requires solid continuous integration. Without automated tests, an autonomous agent fixing one user's edge case will break core functionality for everyone else.

What to Do With This

Map your last 20 support tickets into three buckets: pure answers, code bugs, and balance-sheet actions. Keep the balance-sheet actions behind a one-click manual approval workflow like Stripe Agent Actions, and automate the diagnosis and bug reproduction steps first.