The customization spectrum

Before you fine-tune anything, understand the spectrum of ways to customize an LLM's behavior — and the widely-taught guidance to move up it in order, trying the cheaper, faster options before the expensive ones. Most teams that "need to fine-tune" actually need something earlier on the spectrum.

The spectrum, cheapest to most involved:

1. Prompt engineering — carefully craft the instructions. Zero training, instant iteration, no data pipeline. Astonishingly powerful and the right first move for almost everything.
2. Few-shot (multishot) prompting — include examples of the desired behavior in the prompt. Still no training; often closes the gap when zero-shot prompting isn't enough.
3. RAG (Retrieval-Augmented Generation) — retrieve relevant information at inference time and inject it into the prompt. This is how you give a model knowledge it doesn't have — current facts, your documents, proprietary data.
4. Fine-tuning — actually update the model's weights (or train adapters) on your examples. The most involved: needs data, compute, evaluation, and ongoing maintenance.

(And beyond: distillation, where you train a smaller model to mimic a bigger one.)

Why try the cheaper options first — the vendor consensus. OpenAI, Anthropic, and others explicitly recommend starting with prompt engineering and few-shot, then RAG, and only fine-tuning if those fall short. The reasons:

• Cost and speed — prompting iterates in seconds; fine-tuning takes data, compute, and time.
• No maintenance burden — a prompt or RAG system is far easier to update than a retrained model.
• It preserves the model's general knowledge — fine-tuning risks degrading the base model's broad capabilities (catastrophic forgetting, Module 4).
• It often just works — a great many "we need a custom model" problems are solved by a better prompt or by RAG, at a fraction of the effort.

And even when you do fine-tune, best practice is to fold your best prompt into every training example — the spectrum stages combine, they don't replace each other.

The key reframe: customization is not "prompt OR fine-tune" — it's a spectrum you climb only as far as you need to. Start at prompting, add few-shot, add RAG for knowledge, and reach for fine-tuning only when prompting and RAG genuinely can't deliver what you need (which the next lessons help you recognize). Climbing past what you need adds cost and maintenance for no benefit.

The mindset: LLM customization is a spectrum — prompt engineering → few-shot → RAG → fine-tuning — and the durable guidance is to climb it in order, trying the cheaper, faster, lower-maintenance options first. Prompting and RAG solve a surprising share of "we need a custom model" problems at a fraction of the cost, and they preserve the base model's general knowledge. Reach for fine-tuning only when prompting and RAG genuinely fall short — and even then, keep your best prompt in the training data. Knowing where to stop on the spectrum is the first and most valuable skill in model customization.