Examples are stronger than explanations.
There are limitations to explaining what you want to AI verbally. I understand why showing format and reasoning through examples is more accurate than explanation.
When AI Doesn't Give You What You Want
You put quite a bit of effort into writing a prompt, but the result sometimes differs from your expectations.
You gave it a role, explained the context, and specified the output format. Yet the answer the AI produces seems to be on the right track, but it's not quite what you wanted. The tone is different, the format is slightly off, or the reasoning process is unclear.
In such situations, longer and more detailed explanations can be a solution sometimes, but more often they're not.
Things That Are Hard to Explain in Words
There are things that are difficult to convey through words.
Tone is one of them. Even if you write "professional but not too stiff," the tone the AI thinks of and what you envision are different. Adjectives have no clear standard. The AI selects one of many patterns associated with that adjective from within its training data.
Output format is the same way. "Give it to me in JSON" works, but schemas used only internally at your company or your own classification systems aren't easy to explain in words. The longer the explanation, the more room for interpretation the AI has.
In such cases, you can show examples instead of explaining.
Examples Show the Shape of the Answer
Few-shot is a simple concept. It means directly including examples like "this input produces this output" in your prompt.
Input: "2024-01-15 Meeting Minutes"
Output: {"doc_type": "MOM", "date": "20240115", "priority": "normal"}
Input: "Urgent: Server Failure Report"
Output: {"doc_type": "IR", "date": "today_date", "priority": "critical"}
Input: "Q3 Sales Analysis Report"
Output:
The AI sees this pattern and fills in the last blank. You don't need to explain. You just show.
Examples work because the AI doesn't need to infer the format. Instead of explaining in words what doc_type should be in each case, two completed examples define the standard. This approach is far more stable than text instructions, especially for non-standard output formats, fixing tone and style, and defining classification labels.
Two to five examples are usually appropriate. More than that wastes context and can become noise. Consistency matters more than the number of examples. If the format is mixed even slightly, the AI becomes uncertain.
When Showing the Process of Thought
Format problems can be solved with examples. But reasoning problems are different.
The way AI gets things wrong on complex calculations or multi-step logic problems is unique. Rather than getting the wrong answer, it jumps straight to a plausible-looking one. It skips the intermediate steps and creates the conclusion first.
An LLM generates tokens sequentially, using what it previously generated as context. When the AI actually writes out intermediate steps as tokens, those become the foundation for the next inference. But if it jumps straight to the answer, that foundation is missing.
Chain of Thought leverages this principle. It makes the intermediate steps explicitly output.
# Without steps
Q: 3 apples (1,200 won), 2 pears (2,500 won). Final amount after 10% discount?
A: 8,730 won ← Wrong
# When steps are written out
Q: Calculate the above problem step by step.
A:
Step 1: 3 apples = 1,200 × 3 = 3,600 won
Step 2: 2 pears = 2,500 × 2 = 5,000 won
Step 3: Total = 8,600 won
Step 4: 10% discount = 860 won
Step 5: Final = 7,740 won ← Correct
It's not just about calculations. The same pattern appears in legal interpretation, code debugging, and decision-making with multiple intertwined conditions. When intermediate logic is skipped, a plausible wrong answer results.
You can achieve this effect by simply adding one sentence to your prompt: "Think through this step by step." Or, like Few-shot, you can show the step-by-step reasoning process itself as an example. The model follows that reasoning pattern.
Good Examples and Bad Examples
If examples are wrong, they can backfire.
If Few-shot examples mix inconsistent formats, the AI becomes confused about which standard to follow. If there are too many examples, they eat up your context window and push out what actually matters.
There are things to watch out for with Chain of Thought too. Recent reasoning-specialized models already conduct long reasoning processes internally. When you add "think through this step by step" to such models, you're asking them to do what they're already doing. The output becomes unnecessarily long and costs increase.
The effectiveness of these techniques varies depending on the model you use. The less a model has a reasoning-specialized mode, and the higher the complexity of the task, the more effective these techniques are.
What It Means to Give Examples
Giving examples doesn't mean telling the AI the correct answer.
It means showing what a good answer looks like. Format, tone, reasoning flow — when these are explained in words, room for interpretation arises, but when shown through examples, the standard is fixed.
If explanation gives AI direction, examples give AI shape.