How it works

The value is in the curation, not the model.

You invest in the catalog once and it pays off on every message. The LLM is a safety net for the cases the script can't classify — never the default.

Define exactly which subjects the system is allowed to handle. Anything outside is escalated, never guessed.

For each topic, list every possible question, request, suggestion and option — the full surface of what people ask.

Write the answers in priority tiers: first-order for the common cases, fallbacks for the edges.

When interpretation fails, the model fills in — and you promote the recurring ones back into the script.