Fine-Tuning vs Prompting

Two ways to specialize a general model

Almost nobody outside a handful of large labs trains a foundation model from scratch (see Pretraining: How Models Learn Language). Instead, the practical question most builders face is: given an existing model, how do you get it to do your specific task well? There are two fundamentally different answers.

Prompting

Prompting means giving the model instructions and context as part of the input text, without changing the model itself at all. This ranges from a simple one-line instruction, to detailed system prompts, to "few-shot" prompting where you include a handful of worked examples directly in the prompt so the model can pattern-match to them. Prompting is fast to iterate on, requires no training infrastructure, and works with closed-weight models you can't otherwise modify — which is why it's the default approach for most AI tools and products.

Fine-tuning

Fine-tuning means further training an existing pretrained model's weights on your own (usually much smaller) dataset, nudging its parameters toward your specific task or style. Unlike prompting, this actually changes the model, so the resulting behavior doesn't depend on repeating instructions every time — but it requires real training infrastructure, a genuinely good dataset, and (for closed models) the provider to actually offer fine-tuning access at all.

Which one makes sense

As a rough rule of thumb: prompting is usually the right first move for almost any task, since it's cheap to experiment with and today's frontier models follow instructions well. Fine-tuning tends to make sense when you have a large, consistent dataset representing exactly the behavior you want, a task narrow enough that "teaching" beats "instructing," or a need to run inference cheaply on a smaller fine-tuned model instead of a larger general-purpose one. Many production systems end up using both together — a fine-tuned base with careful prompting on top.