Step by Step
P
Prompting — fast, cheap, no training required
Prompting means writing better instructions, adding few-shot examples, and structuring your query more effectively — fast and cheap, requiring no training at all, and sufficient for the vast majority of tasks.
Example: improving a prompt by adding three clear examples of the desired output format, rather than retraining the model.
F
Fine-tuning — updating the model's weights on your data
Fine-tuning updates the model's actual weights using your specific task data, needed when the task requires a consistent style or format that prompting alone struggles to achieve, when you have thousands of task-specific examples available, or when you need performance beyond what prompting can achieve.
Example: fine-tuning a model on thousands of examples of a very specific, consistent document format that prompting alone couldn't reliably replicate.
⭐
The rule of thumb — always try prompting first
The standard guidance is to always try prompting first, since it's fast and cheap, and to only move to fine-tuning if prompting genuinely hits a performance ceiling that better prompting can't overcome.
Example: spending time refining prompts with better instructions and few-shot examples first, and only considering fine-tuning after confirming that prompting improvements have plateaued.
Applied Walkthrough
1
A team needs a model to consistently follow a very specific, complex document formatting style, and their initial prompting attempts aren't achieving reliable enough consistency.
2
Following the standard rule of thumb, they first invest more effort into improving their prompts — adding more few-shot examples, clearer instructions, and better structure.
3
If this refined prompting still doesn't achieve the necessary consistency, and they have thousands of examples of the desired format available, fine-tuning becomes the appropriate next step.
4
This progression — always trying prompting first, only fine-tuning once prompting genuinely hits a ceiling — reflects the standard, cost-effective approach to customizing an LLM for a specific task.
Exam Application
Exams test whether you understand the specific conditions favoring fine-tuning (need for consistent style/format, having thousands of task-specific examples, hitting a performance ceiling with prompting) versus when prompting alone is sufficient, and whether you know the standard rule of thumb: always try prompting first.
⚠ Common Trap
The most common trap is jumping straight to fine-tuning without first exhausting prompting-based improvements. Fine-tuning is more expensive and time-consuming than prompting, and the standard guidance is to only pursue it once prompting has genuinely hit a ceiling that better instructions and examples can't overcome.
✓ Quick Self-Check
1. What are the main advantages of prompting over fine-tuning?
It's fast, cheap, and requires no training at all.
Tap to reveal / hide
2. Name one condition that would justify fine-tuning instead of just prompting.
The task requires a consistent style/format that prompting alone struggles to achieve, having thousands of task-specific examples, or needing performance beyond what prompting can achieve (any one).
Tap to reveal / hide
3. What is the standard rule of thumb for choosing between the two?
Always try prompting first; only fine-tune if prompting hits a ceiling.
Tap to reveal / hide
4. Does fine-tuning update the model's weights, or just its input prompt?
Fine-tuning updates the model's actual weights, unlike prompting which only changes the input given to the unchanged model.
Tap to reveal / hide
5. Is fine-tuning generally faster and cheaper than prompting, or the reverse?
The reverse — fine-tuning is generally more expensive and time-consuming than prompting.
Tap to reveal / hide