Stop Fine-Tuning When a Prompt Will Do

The task was classifying support tickets by intent. I assumed I needed a custom model. I built one. It was fine. Then I tried a careful prompt with examples on a frontier model and it matched the accuracy.

Fine-tuning bought me lower latency and a smaller bill, but only after I'd proven the prompt version worked. Starting with the model would have saved the weekend.

Reach for fine-tuning when prompts hit a wall on volume or cost, not as a first move. Most 'we need a custom model' claims fail this test.

Prompt first. Specialize later, and only if the bill or the latency forces your hand.