Fine-tune or prompt

Fine-tuning or prompt: which one for my case?

Prompting, nine times out of ten. Fine-tuning a model sounds professional, but it means data to prepare, training to maintain and a cost that comes back with every change to the model. Before going there, I make sure the need is real and that prompting has genuinely hit its limits.

What does prompting already do very well?

Output format, tone, business rules, examples: the vast majority of needs are handled by good instructions and a few well-chosen examples. It is immediate, adjustable, and it creates no technical debt.

When is fine-tuning genuinely worth it?

In a minority of cases, when prompting hits a ceiling despite every effort. Concretely, three situations justify it.

• A very specific style or format, required at very high volume.
• A latency or a cost to compress on a task repeated millions of times.
• A domain where prompting hits a ceiling despite every effort.

My rule: prompt first, measure, and only fine-tune if the numbers demand it. Nine times out of ten, they don't.