Decision heuristic 2025-2026: (1) prompt engineering first: try structures like CoT, ReAct, tool-use scaffolding with the base model. If that hits target quality, stop. (2) RAG: when the task needs fresh, document-specific, or proprietary knowledge that the base model doesn't have. RAG is preferred over fine-tuning for knowledge updates because you can swap documents without retraining, you get provenance, and it's cheaper. (3) Fine-tuning: when you need to change behavior (style, format, tool-use schema, output structure, domain jargon compression) or compress a costly prompt into learned weights. Concretely: I fine-tune when (a) I need the model to consistently do something in a specific format thousands of times per minute (latency/cost), (b) I have 10k+ high-quality supervised examples, (c) I'm online-learning from production feedback. Senior nuance: hybrid is common (RAG for grounding + light fine-tuning for persona/tool schema) and evaluation discipline separates the good practitioners from those who fine-tune prematurely.