Prompts teach patterns, not facts

Look at almost any prompt that's been in production a while and you'll find it quietly rotting — stuffed with facts. "Gmail is supported. The crypto tool takes a symbol. Connected services are prefixed with svc_. Respond in the user's language." Every one of those is a fact frozen into a string, and every one of them goes stale the moment the world moves.

Across every version of my workflow engine, the same move kept paying off: drag the facts out of the prompt and into a runtime signal — a function call, the request context, a catalog schema. The prompt stops asserting what's true and starts teaching how to find out. It teaches pattern; the truth arrives live.

The payoff compounds in three directions, and you feel each one within a release or two:

The second is honesty under change. A hard-coded "Gmail is supported" is a lie waiting to happen the day Gmail's auth breaks. A runtime health signal that says "this source is degraded" lets the model route around it and tell the user the truth. The prompt can't know the weather; the context can.

The third is generality — the quiet one. The moment you stop hard-coding zh/en branches and let a detected signal carry the language, the same prompt works for Japanese, Arabic, Korean, languages you never tested. You didn't add support; you stopped removing it. Over-fitting a prompt to the cases in front of you is how you accidentally forbid the cases you haven't seen.

There's a deeper reason this is the right shape for models specifically. A model is a pattern matcher with a context window, not a database with a schema. Ask it to memorize facts and you're using the most expensive, least reliable storage ever built. Ask it to read signals and reason over them and you're using the thing it's actually good at. Keep the facts in the systems that own them. Let the prompt teach the model how to ask.