Four ways to shape model behaviour at wildly different costs, and how to choose. Most problems taken to fine-tuning are context problems — and a fine-tune expires with its base model.
- If the model does not know your product, your data or your policies, supply them — see context engineering.
Four ways to make an AI behave how you want, from cheapest to most expensive:
Ask better. Then: give it the material instead of describing it. Then: save your instructions so you do not retype them. Last: fine-tuning, which means retraining the model — weeks of work that stops being useful when the model is replaced.
Almost everyone jumps to the last one. Almost everyone should have stopped at the second.
Four ways to make a model behave a particular way, at very different costs. Most people reach for the most expensive one third.
The ladder, cheapest first
How to choose
If it is a one-off, it is a prompt
Anything you will do once does not justify a reusable artefact. The reusable artefact costs more to write than the task costs to do.
If the model lacks information, it is context — never fine-tuning
The most common expensive mistake. Fine-tuning teaches behaviour, not facts. If the model does not know your product, your data or your policies, supply them — see context engineering.
Fine-tuning on facts produces a model that is confidently wrong in your house style.
If you repeat the same shaping, it is a skill
The same tone, the same checks, the same format, every time, across sessions. That is what a skill is for. Read the supply-chain page before installing anyone else's.
Fine-tune only for form at volume
A consistent output shape, at a scale where prompt overhead is a real cost, in a domain that will not change. Three conditions, all required.
And the cost nobody prices: a fine-tune is pinned to its base model. When that model is retired — as Sora was, with an API shutdown scheduled for 24 September 2026 — the work goes with it.
The honest summary
Most problems people take to fine-tuning are context problems, and most problems people take to prompting are also context problems.
The ladder is worth climbing in order because each rung usually solves it. The reason people skip to the top is that fine-tuning sounds like the serious answer, and supplying better context sounds like admitting you had not.