We all know LLMs are brilliant at coding and writing. But if you ask an LLM to guide a foreigner through the physical world (e.g., navigating a Chinese airport, using an Alipay QR code, or surviving a Chinese ER), it generates "perfectly formatted garbage."

Why? Because LLMs optimize for fluency, not physical feasibility.

Recently, we built GoChinaTrip / SSAP. I fed the AI a real-world scenario: "A foreigner is lost. The DiDi driver is calling. What should they do?"
The AI suggested: "Show a translated screen to the driver."

This is physically impossible. The driver is driving. The foreigner is standing on the curb.

To fix this, we wrote SSAP Card 00: The AI Mandatory Constraint Protocol. We stopped treating the LLM as a "writer" and started treating it as a "constrained parser."

Here is the exact open-source Markdown structure we use to force the AI to respect physical reality:

SURVIVAL ACTION CARD 00: Anti-Sycophancy & Physical Reality

Rule 0 — Reality Before Requirements

AI must NOT extract text blindly. It must identify:

  1. The Environment: Where is the user physically located?
  2. The Primary Contradiction: What is the deadlock? (e.g., The app says 'Paid', but the register didn't ring).
  3. The Physical Channel: How can the user communicate? (Voice? Screen? In-app chat?)

Rule 1 — No Unauthorized Amputation (禁止擅自阉割)

Do NOT delete user-provided physical details (like "look for the blue sign") to make the text shorter or more "professional."

Rule 2 — Physical Feasibility Check

Before outputting any instruction, run this check:

  • Can this action be performed at 2 AM in the rain with heavy luggage?
  • Is the communication channel actually accessible? (e.g., NEVER tell a user to show a phone screen to a driver who is navigating traffic). By enforcing these strict JSON/Markdown schemas, we completely eliminated the "hallucinated convenience" of LLMs. We forced the AI to output pure physical truth:

"Find the designated '网约车' zone BEFORE opening the app."

"Go to the Triage Desk first, pay for tests BEFORE getting them."

"Give the cashier this exact Chinese script if they try to double-charge you."

The AI provides the data. The humans (in this case, my 77-year-old mother and I, testing these routes in the rain) provide the pain.

If you are building AI agents for the physical world, stop teaching them to write. Start teaching them to respect physical constraints.

Check out the full SSAP repository on GitHub: [Link]
See the interactive survival cards at GoChinaTrip.app