Last week I built a prompt framework, sent it to an AI connected with travel data capability, and it completed the entire hotel price comparison workflow for me.
Bottom line: With this prompt framework, you just tell the AI your destination, check-in date, and price range. It handles city-wide hotel filtering → 5-candidate horizontal comparison → actionable booking recommendation.
It even has a proactive capability — when your target hotel drops in price, the AI pushes a notification.
The Prompt Framework: Search City-Wide, Compare 5, Then Decide
This framework has three phases. You don't start comparing — you first find what's worth comparing.
Phase 1: Input Conditions, Search City-Wide
`Here are my travel needs:
- Destination: {city / district}
- Check-in date: {checkin_date}
- Check-out date: {checkout_date}
- Budget per night: {budget}
- Other preferences: {star_rating / breakfast / free cancellation / near metro}
Please filter from the global hotel database for candidates matching the above:
1. Rank by "value for money," show me 20 best candidates
2. For each: hotel name, rating, location, total price with tax, deviation from budget
3. From these 20, select the best 5 for deep comparison`
This phase solves "where to look." Not scattering wide, but filtering a candidate set from the global hotel pool, then selecting 5 for the next phase.
Phase 2: 5 Candidates, 4 Dimensions Deep Comparison
Dimension 1: Price Trend Analysis
`My check-in date is {checkin_date}, check-out is {checkout_date}.
Analyze the 7-day price trend for each of these 5 hotels:
- When does each hotel's lowest point appear?
- What's the difference between the lowest price and current price?
- If I flexibly adjust check-in date (±1 day), is there a cheaper option?
Provide a "recommended check-in date" for each hotel.`
Hotel prices are more directly supply-demand driven than flights — weekends up, holidays up, conference periods up, then back down. The AI needs to find each hotel's individual low point, not generically say "prices are okay recently."
Dimension 2: Cancellation Policy Assessment
`For these 5 hotels, query and compare their cancellation policies:
- What is each hotel's latest free cancellation time?
- If I book now, cancel and rebook, within what price difference is it worth waiting?
- If the price hasn't dropped by the free cancellation deadline, what's the actual risk?
Provide a "breakeven value" for each hotel: above what price drop is waiting one day worth it?`
This is the most overlooked dimension in hotel scenarios. Flight changes have costs, but hotels within the free cancellation window can be cancelled and rebooked with near-zero loss. The key calculation: is the money saved by waiting one day worth gambling the "loss of free cancellation eligibility"?
Dimension 3: Competitor Price Comparison
`For these 5 hotels, are there lower rates on major global booking channels?
- Which channel has the lowest price for each hotel?
- What's the difference between the lowest rate and current quote?
- If a lower channel exists, can the original hotel match it?
Also analyze: are there better alternatives at similar or lower prices
in the candidate pool? Give Top 2 alternatives.`
Among 2M+ global hotels, each property's price across channels can vary by 20% to 2× — not information asymmetry, but supply chain complexity. The AI needs to scan beyond these 5 for "same rating but cheaper" or "better value at similar price" gems.
Dimension 4: Final Recommendation (with Action Instructions)
`Based on all four dimensions (price trend / cancellation / competitor pricing /
candidate supplements), provide the final decision:
1. If I must book today, which hotel? Why?
2. If worth waiting, until when? Will free cancellation still apply?
3. Which booking channel? Can price be adjusted later?
4. Final recommendation: hotel name, channel, check-in date, total price,
free cancellation deadline.`
Good AI doesn't just give information — it gives executable action instructions.
Real Results
I ran this framework on a real scenario: Osaka, a weekend in July, 3 nights, budget under $120/night.
Input: destination, check-in date, price range. That's it.
The AI completed all three phases:
Phase 1: Filtered 20 candidates from Osaka Namba area, ranked by value, showing rating, location, and total price for each.
Phase 2: Selected 5 for deep comparison — price trends, cancellation policies, competitor pricing. Found one hotel 12% cheaper through comparison, and two with free cancellation 48 hours before check-in but prices still trending down.
Phase 3: Final decision —
- Option A: Book now, Hotel B, via Booking.com, 3 nights total $X, breakfast included, free cancellation until 48 hours before check-in.
- Option B: Wait 2 days, Hotel C has ~60% probability of price drop, could save ~$Y, but Hotel B's optimal cancellation window will have passed.
- Option C: Switch to Hotel D, 10 minutes' walk away, same conditions, $Z cheaper, rating difference within acceptable range.
Three options, each with price data backing, cancellation policy details, and clear action instructions. Not hallucinated data — directly actionable results.
The Real Barrier: Where Does the Data Come From?
The prompt framework itself isn't complex. The actual problem is the data source.
I run this framework using RollingGo Hotel MCP — real-time data from 110K+ direct-contracted hotels. Not scraped (I tried scraping — two weeks later, OTA anti-scraping kicked in; you think you're getting real-time prices, but you're getting what the platform wants you to see). Not cached. 2M+ global hotels, 500+ suppliers — this scale is the prerequisite for running the entire "city-wide filter → candidate comparison → competitor cross-check" chain.
If the AI tells you "$150 tonight" but that's a 3-day-old cache and the room is actually sold out — the prompt framework doesn't matter how well-crafted it is.
Advanced Capability: Proactive Price Drop Alerts
MCP solves the query problem: you ask, the AI queries, returns results.
But there's a scenario MCP doesn't cover: proactively notifying users when prices drop.
You're eyeing a hotel, waiting for the price to drop before booking, but you can't monitor 24/7. RollingGo also offers a price monitoring Skill — install it, and when your target hotel hits your target price, it pushes a notification.
MCP is "you come to ask." Skill is "it comes to tell you." Two different problems, but they can be combined: use Skill to monitor prices, then use MCP to confirm inventory and book when the target price hits.
Config:
- Apply for API key at global.rollinggo.store
- Configure MCP in Claude Code or Cursor:
`{
"mcpServers": {
"rollinggo-hotel": {
"type": "streamable-http",
"url": "https://mcp.rollinggo.ai/mcp",
"headers": {
"Authorization": "Bearer YOUR_API_KEY"
}
}
}
}`
- Restart client, tools appear. Use the prompt framework in your agent.
Pitfall: The space between "Bearer" and the key is easily missed. If type is set to http but your client only supports streamable-http, tools won't appear — check these two items first.
Summary
Three phases, three layers of prompts:
- Search city-wide — Filter 20 candidates by your conditions, rank by value
- Compare 5 — Deep-dive 4 dimensions (price trend / cancellation / competitor pricing / alternatives)
- Decide — Give an executable booking instruction
The framework is easy to copy. What's hard is the data source. Without real-time inventory, the AI always gives cached reference prices — not tonight's transaction price.
Choosing the right data source matters more than writing the perfect prompt.
RollingGo Hotel MCP — fully open-source, real-time hotel data for AI agent developers. 2M+ hotels, 110K+ direct-contracted inventory, 500+ global suppliers.
Get your free API key: https://global.rollinggo.store/
MCP endpoint: https://mcp.rollinggo.ai/mcp
5-Minute Quick Start: https://global.rollinggo.store/docs/mcp-docs/quick-start
Top comments (0)