Prompt first when the requirement is still moving
A prompt is easier to edit than a training set. If product terms, tone, fields, or approval rules are changing every week, fine-tuning will make a moving target expensive to maintain.
Prompt engineering is usually enough for
- A clear task with a small number of examples.
- Output formatting and schemas.
- Role, audience, tone, and length constraints.
- Adding a current source through retrieval.
- Early prototypes where you are still learning what good looks like.
Fine-tuning becomes interesting when
The same failure repeats across a stable task, you have enough approved examples, and the improvement can be measured. Fine-tuning may help with classification, consistent transformations, or a stable style, but it will not automatically add current knowledge.
A practical order of operations
| Try first | Why |
|---|---|
| Clarify the task | A vague requirement cannot be trained reliably |
| Add examples | Few-shot examples often fix format and style issues |
| Add retrieval or tools | Use the right source instead of teaching changing facts |
| Build an evaluation set | Measure the failure rather than trusting anecdotes |
| Fine-tune | Only when the stable behavior justifies maintenance |