The Customization Spectrum

Five techniques, one spectrum

Before you fine-tune anything, place the problem on the spectrum. Each technique trades effort for control, and most projects pick the wrong rung — usually one too high.

Technique	Cost	Effort	When it shines
Prompt engineering	Free	Minutes	~90% of practical use cases. Always start here.
Few-shot in context	Free	Minutes	Consistent output format, structured replies.
RAG	Low	Days	Model needs your data or documents that change.
Fine-tuning	Medium	Days–weeks	You need a behavior — style, format, refusal — to be reliable.
Pre-training from scratch	Very high	Months	You are building a foundation model. Almost no team should be here.

RAG vs fine-tuning, one sentence each

RAG changes what the model knows at query time by injecting fresh context. Fine-tuning changes how the model behaves by adjusting its weights. They are not substitutes — most production systems use both.

The honest decision rule

Before reaching for fine-tuning, force yourself to answer: "What did the strongest prompt I could write get me, and where exactly did it fall short?" If you cannot describe the gap precisely, you are not ready to train. Fine-tuning will not rescue an undefined goal.

Code

Baseline that has to exist before you train·python

from openai import OpenAI

client = OpenAI()
SYSTEM = """You are a JSON-only product-review classifier.
Output: {"sentiment": "positive|negative|mixed", "confidence": 0..1}.
No prose, no markdown."""

def baseline(review: str) -> str:
    r = client.chat.completions.create(
        model="gpt-4.1-mini",
        messages=[
            {"role": "system", "content": SYSTEM},
            {"role": "user", "content": review},
        ],
        response_format={"type": "json_object"},
        temperature=0,
    )
    return r.choices[0].message.content

# Run this against 200 labeled reviews FIRST.
# If it already hits 95%+ format compliance and your tolerated metric,
# do NOT fine-tune. The baseline is the gate.

Exercise

Pick a real task. Write the strongest single prompt you can for it, then evaluate against 50 representative inputs. Record format compliance and task accuracy. Only after that, write a one-paragraph fine-tuning brief: target metric, current metric, the specific gap, and what kind of training data would close it.