Training, fine-tuning and prompting are three different tools
Pre-training builds general capability, fine-tuning adjusts behaviour and format, and prompting supplies task-specific context at request time. Fine-tuning is poor at teaching facts, which is the job most teams mistakenly hire it for; retrieval or a longer prompt is almost always the right answer there.
Three words get used as if they were points on one scale, from cheap to expensive. They are not. They solve different problems, and choosing by budget rather than by problem is how teams spend a quarter fine-tuning their way to something a prompt already did.
Pre-training: not your problem
This is where general capability comes from — vast corpora, enormous compute, months of work. It is the reason a model can write at all.
For almost every team this is background, not a decision. You choose a model that was pre-trained well and move on.
Fine-tuning: teaching behaviour
Fine-tuning continues training on your examples, nudging the model towards a particular way of responding. It is very good at exactly that: a consistent output shape, a house voice, a domain-specific label set, a terse style where the base model is verbose.
What it is not good at is facts. Knowledge that arrives through fine-tuning is diffuse and unreliable — the model absorbs the shape of your documents more readily than their content, and it cannot tell you when it is unsure.
The cost nobody budgets for
The training run is the small part. The lasting costs are curating examples, re-running when the base model updates, and losing the ability to switch providers cheaply. A fine-tune is a commitment, not a setting.
Prompting: supplying context
Everything you put in the request — instructions, examples, retrieved documents, the user's own words — is prompting. It is the only one of the three that operates at request time, which gives it a property the others cannot have: it is current.
A price that changed this morning, a document edited ten minutes ago, the specific row this user is looking at — all of it can be in the context. None of it can be in the weights without another training run.
How to choose
Start with the question "is this about what the model knows, or how it behaves?" Knowledge goes in the context, through retrieval if there is too much of it to send. Behaviour goes in the instructions, and only moves to fine-tuning when instructions have been tried properly and the inconsistency is still costing you.
The order matters because the cheap options are also the fast ones. A prompt change ships in minutes and can be reverted as quickly. Exhaust that before you buy something permanent.
Frequently asked questions
Can I fine-tune a model on our documentation so it knows our product?
You can, and it will usually disappoint. The model picks up the tone and vocabulary but not reliable recall of specifics, and anything that changes leaves you with stale weights. Retrieval handles this better and updates instantly.
When is fine-tuning clearly the right call?
When you need a narrow behaviour repeated consistently and cheaply — a fixed output format, a house style, a classification task with an idiosyncratic label set — and you have a few hundred good examples of it.
