Treasure Maps Without AI: When the LLM Is the Wrong Answer

Our NPCs can give players hints about treasure — an abandoned wreck or, if you're on good terms, a hidden chest. When we planned the feature, one solution seemed obvious: we have an AI backend, so why not let the NPCs have the language model phrase the hints! We decided against it — and the reason makes a nice lesson.

The problem with the LLM: Our NPC dialogue from the AI backend is cached for cost reasons. For small talk that's perfect — nobody notices whether the trader delivers the same line today or yesterday. But a treasure hint contains facts: a direction, a distance, a location. A cached hint would point to yesterday's chest. And even uncached, language models tend to creatively "improve" coordinates. A hint that leads you astray is worse than no hint at all.

The solution: deterministic. The hints are now built by hand, the classic way. The server knows the real position, calculates the direction, and composes a spoken line from templates (German and English). At the same time, the location is revealed as a point on the map for all players in the world. A wreck that's already been looted automatically drops out of the pool. No model, no hallucination, no stale caches — the hint is simply correct.

The rule we drew from this: For us, LLMs handle atmosphere — personality, small talk, narrative tone. As soon as a statement has to be factually correct (places, numbers, game rules), deterministic code is the right technology. Not every problem that looks like an AI problem is one.