Fine-tuning vs prompting: a decision you can defend
Fine-tuning is rarely the first answer and occasionally the only one. A short decision procedure for choosing between prompt engineering, retrieval, and training.
The question arrives early in most projects and gets answered by preference rather than evidence. Here is the order we actually work through.
Start with the prompt, but properly
Most "the model can't do this" conclusions come from a prompt with no examples, no output schema, and no failure instructions. Few-shot examples drawn from real cases usually close the gap on their own.
If it needs facts, that is retrieval
Knowledge the model lacks is a retrieval problem, not a training one. Fine-tuning teaches behaviour and format; it is an expensive and lossy way to teach facts that change.
Fine-tune for form, tone, and cost
Training earns its keep when you need a consistent house style, a rigid output structure, or the same quality from a smaller and cheaper model. Those are real wins — they are just narrower than people expect.
Count the total cost
A fine-tune is not a one-off. It is a dataset to maintain, a pipeline to run, evaluations to keep current, and a retraining commitment every time the base model moves. Take it on deliberately.
