Fine-tuning is good at teaching a model how to behave, not what to know. It shifts form, style and task habits, which you can show in examples but struggle to describe in a prompt.
The usual sweet spots:
Consistent format, tone or house style. Examples beat a long, brittle style prompt.
One narrow, repeated task, such as classifying, extracting or routing. The model only needs that job done well.
Cost and latency. A small tuned model can match a large prompted one.
Unusual output shapes or tool-call patterns. They are easier to demonstrate than to explain.
It also shortens prompts. Behavior baked into the weights does not need repeating on every call.
Rewriting in plainer words…
This answer doesn't lend itself to a diagram - it reads best . No credits were charged.
What would you try before fine-tuning for a house style, and what tells you it's not enough?Start with instructions and few-shot examples. Move on when the prompt keeps growing, output still drifts, or per-call cost and latency hurt.
What you can say
Fine-tuning is good at teaching a model how to behave, not what to know.
It shifts form, style and task habits, the things you can show in examples but struggle to describe in a prompt.
A consistent format, tone or house style is the classic case, because a set of examples beats a long, brittle style prompt.
It also fits one narrow task you run over and over, like classifying, extracting or routing, and odd output shapes or tool-call patterns that are easier to demonstrate than explain.
Then there's cost and latency, since a small tuned model can match a large prompted one, and behavior baked into the weights doesn't need repeating in every prompt.
Where I wouldn't use it is for new facts, because it teaches behavior, not knowledge, so missing knowledge belongs in the prompt or in retrieval.
Weak answers to avoid
Says fine-tuning adds domain knowledgeThis is the core misconception. Tuning shapes behavior, and facts it absorbs are unreliable and hard to update. Say knowledge goes in the prompt or retrieval.
Reaches for fine-tuning before trying the promptTuning costs data, training runs and upkeep. Say you'd start with instructions and few-shot examples, and tune when the prompt gets long, brittle or expensive.
Lists use cases with no reason behind themA bare list sounds memorized. Tie each case to the rule that it's behavior you can show in examples but can't describe well in a prompt.
Leaves out the cost and latency caseSwapping a large prompted model for a small tuned one is a common production reason to tune. Skipping it misses the business case interviewers often want.
Interview lens
Likely follow-ups, what you can say, and the weak answers to avoid.
