keep the model in its lane

Both of my LLM projects ended up with the same rule: the model only does the work that actually needs a model. Everything else is code.

In the tone translator, the model translates and explains nuance. Detecting the direction (JP→EN vs EN→JP) is a Unicode regex, because a model asked to figure out the direction will occasionally echo Japanese back as Japanese. In riichi, the daily puzzle’s answer comes from a shanten library, so it can’t be wrong; the model only writes the explanation. The AI opponents aren’t a model at all, just hand-written rules. They act dozens of times a game, and a model there would cost real money for zero teaching value.

A model is a fuzzy-judgment engine, not a source of truth. If something is computable, compute it: cheaper, faster, testable. A lot of the job is knowing what not to hand the model.