Course Appendix & Quick-Reference Card

Keep this page open while you work. It condenses the whole course into things you can copy.

1. The pattern

Every strong instruction contains Goal · Scope · Limit · Why:

• Goal — what you want done.
• Scope — which dates, room types, staff, or guests.
• Limit — the hard boundary (price, count, budget, deadline).
• Why — the reason, so the AI can learn the preference.

The Goal/Scope/Limit make a recommendation correct; the Why is what trains the system.

2. Gold-standard prompts, one per bee

Pricing (WB1): "Deluxe rooms are slow for Thursday. You may discount to fill them but never below 3,000, and stop once we hit 85% occupancy — I'd rather protect rate than sell out cheap." (Amounts are always in your hotel's base currency.)

Ops / dispatch (WB2): "Prioritize checkout-ready rooms for the 2 pm arrivals; inspect Room 204 first for an early VIP, and flag anything needing maintenance instead of just cleaning."

Marketing (WB3): "Write a short, warm SMS for repeat Dashain guests: 15% off a 3-night stay booked before Ghatasthapana, one clear booking link."

Reviews (WB4): "Draft a warm reply to the 4-star Booking.com review about breakfast timing. If any review looks fake or targets a competitor, flag it for dispute instead."

Complaints (WB7): "AC not cooling in Room 302; guest checked in an hour ago and is uncomfortable — needs a fix or a room move tonight."

3. Glossary

• Bee — a specialized AI assistant for one domain (WB1 pricing, WB2 dispatch, WB3 marketing, WB4 reviews, WB7 complaints).
• Co-Pilot — default mode. The AI recommends via cards; you approve, edit, or reject before anything happens.
• Auto-Pilot — earned, invite-only mode where a qualified stream acts on its own within your approved limits, logging every action. Requires ≥90% agreement over ≥100 decisions across ≥60 days with zero breaches.
• Confirmed Policy — a remembered rule the AI forms from your repeated edits and approvals, which you confirmed on the Policies page. Badged Enforced when it carries a machine-checkable rule (the write layer refuses anything that breaks it) or Advisory when it is free text the AI follows in its recommendations. Review, confirm, veto or retire it on that page; policies are not edited in place.
• Guardrail — a founder- or owner-set safety limit (e.g. the ±15% single-step price band). A breach demotes a stream out of Auto-Pilot immediately.
• Kill Switch — the master brake that freezes all background AI activity at once.
• Action Ledger — the permanent, append-only record of every AI action.
• Folio — a guest's running bill for their stay.

4. Auditing: "why did the AI do that?"

To check why a rate changed or why a room was assigned:

1. Open the Action Ledger in settings.
2. Search by room number, guest name, or date.
3. Each entry shows the acting bee, the timestamp, the governing policy (if on Auto-Pilot), and the reasoning behind the decision.

Because the ledger is append-only, it is a complete and tamper-evident history — the source of truth whenever you or an auditor asks what the AI did and why.