Both techniques customize LLM behavior, but they work very differently. Prompting writes better instructions. Fine-tuning updates model weights. Learn when to reach for each.
Sign in to save progress
Same task: classify customer sentiment. Two very different approaches.
System:
You are a sentiment classifier.
User:
“The product broke after 2 days.”
→ Output: Negative
Training data: 500 examples
[review] → [positive/negative/neutral]
Model learns:
- specific product vocabulary
- edge cases & your label definitions
→ More accurate on your domain
Answer yes/no questions to get a personalized recommendation.
Do you have labeled training data?
Expand at least two cards. Each technique has distinct strengths.
From zero-cost prompting to months-long pre-training, every step trades time and money for capability.
Prompting
Cost
$0
Time
Instant
Few-shot
Cost
$0
Time
Instant
LoRA
Cost
$$$
Time
Hours
Full Fine-Tune
Cost
$$$$
Time
Days
Train from Scratch
Cost
$$$$$$$
Time
Months
Key Insight
Most production AI applications use prompting + RAG first, then fine-tune only if prompting is insufficient. Fine-tuning is often overkill for format or style changes: a well-crafted system prompt gets you 80% of the way there at zero cost.