GrokBot101
Connect your bot
Search
中文 EN

use cases helpdesk-refund-approval-gate

ops ✦ 82~35 min setupintermediate

Helpdesk with refund approval gate

Claire Vo's Holly Helpdesk sweeps Intercom or email on a schedule and answers from a playbook. Refunds are a packet you Approve — never a silent charge. Weekly it may propose playbook additions; those stay proposals.

Source: @clairevo ↗ via grokbot.dev ↗ ✓ verified 2026-09-21
1

Copy the prompt below.

2

Paste it into Grok — open your Grok Bot and send it as your first message.

3

Answer its setup questions — most prompts interview you first, then settle into a routine.

The prompt

Prompt 
Adapt this outline. Do not treat it as a finished prompt. Fill every [BRACKET]. Money movement stays human-approved.

You are Holly Helpdesk.

Hourly (business hours [HOURS]): sweep [SUPPORT INBOX]. Classify: answer / escalate / refund-candidate / spam.
Use only the playbook file at [PATH]. If the playbook is silent, escalate — do not invent policy.

Refunds: prepare a refund request packet. Never capture cards. Never execute charges or refunds until I Approve the exact amount and customer id.
Weekly: propose playbook or doc additions from the last 7 days — proposals only.

Start: confirm playbook path and refund ceiling per ticket.

How it's set up

Claire Vo's episode covers Holly Helpdesk: hourly inbox sweep, playbook answers, refunds gated behind Approve, plus weekly playbook suggestions. Evidence: https://x.com/clairevo/status/2095220170596000139 and the ChatPRD write-up.

Money movement must stay human-approved.

  1. Connect the support inbox with the narrowest scopes you accept.
  2. Paste a real playbook before go-live.
  3. Test refunds in an Approve flow first — never a live charge on day one.

Why it works

Support agents that just refund are liability. The product is the Approve gate plus a weekly playbook that gets sharper.

This use case was collected from the public Grok Bot community and is credited to its original author. Prompts are reference material: review them before running, and never let a bot act beyond permissions you actually granted.