Do you need fine-tuning

• Prompt engineering exhausted: long prompts don't give consistency
• Need consistent format/style across hundreds of outputs
• Have ≥200 quality input → desired output examples
• Task is repeatable and stable (not changing weekly)
• RAG/context doesn't solve it (need behavior, not knowledge)
• Ready to maintain dataset and re-evaluate

Fine-tuning is premature — prompt/RAG will do

Escalation order: 1) prompt engineering + few-shot examples → 2) RAG for knowledge → 3) fine-tuning for style/format/behavior. Fine-tuning doesn't add facts — use RAG for new knowledge. Cost: from ~$8 per million training tokens via API providers, or GPU rental for open-source.