Write custom GPT instructions that answer only from your help center

About this exercise

You're setting up a custom GPT that drafts replies to customer questions; the test runs your instructions on three real questions. It must answer only from the help articles provided (using the newest one when two disagree), escalate anything they don't cover instead of guessing, keep each reply under 40 words and end each reply with "- Pebble Support". Output exactly 4 numbered plain-text lines: one action per question in order, then DONE.

Common mistakes when you ask AI for this

• Not showing the log format. Left alone, the AI writes a friendly report with headings instead of a checkable action log. Try: Show it literally: "Format every line like 1. ACTION: tool(args), then ASK_HUMAN: and DONE: lines. Plain text."
• Allowing outside answers. The AI makes up a friendly price-match policy because it sounds helpful. Try: Add: "Answer only from the articles. If they don't cover it, escalate."
• No rule for conflicting articles. The AI uses the first article it finds and quotes the old Canada shipping time. Try: Add: "If two articles disagree, use the newest one."
• No length or sign-off. The AI writes long, warm replies with its own sign-off. Try: Say: "Each reply under 30 words, ending with - Pebble Support."

What a strong prompt includes

• Limits answers to the articles
• Says to escalate gaps
• Uses the newest article
• Sets the sign-off
• Sets a length

Practice this now ↑

Related exercises

• Accept a job offer and confirm the exact terms (Writing AI · Job seekers)
• Allocate shared rent by headcount using the right roster (Finance AI · Accountants)
• Announce a seasonal offer with its real terms and no fake scarcity (Sales AI · Small business owners)
• Answer a 'where is my order' email with the real new date (Writing AI · Customer support teams)
• Answer a big discount ask with two priced options (Writing AI · Sales and customer success)