FEW-SHOT PROMPTING

Show, don't tell. Give Claude 2-3 examples of what you want and it nails the next one. The single most powerful prompting technique most people never use. Examples beat instructions every time.

THE PROBLEM

You explain in detail what you want. Claude misses the mark. You explain again. Still off.

The reason: telling Claude "write in a casual punchy tone with rhetorical questions and one-word sentences" is harder for it to parse than just SHOWING IT 2 examples.

THE PATTERN

I'm going to give you 2-3 examples of what I want, then ask you to do one in the same style.

Example 1:
[paste full example]

Example 2:
[paste full example]

Example 3:
[paste full example]

Now do one for: [your specific input]

That's few-shot prompting. The "shots" are the examples. Few = 2-5 of them.

WHY IT BEATS INSTRUCTIONS

Instructions: "Write a punchy Instagram caption with a hook in the first line, no hashtags, casual tone, sentence fragments OK."

Few-shot: [3 actual captions you've written that you love]

The few-shot version captures rhythm, vocabulary, references, edge — things you can't easily articulate in instructions.

REAL EXAMPLE

Bad (instructions only):

Write me an Instagram caption about meal prep in a casual mom voice.

[Generic output]

Good (few-shot):

Here are 3 Instagram captions I've written that I love:

1. "Sunday meal prep: 4 dinners, 1 hour, $32 at Trader Joe's. Recipe in comments. Yes I'm tired. No I'm not making toast for the kids again."

2. "5 things I'm not buying this week: organic strawberries ($8?!), name-brand cereal, anything 'family-size' that's actually small, the candle, the candle's friend."

3. "If you've made it past 9pm with all kids alive, you're winning. The dishes can wait. Show me your kitchen, no tidying."

Now write one for: a Tuesday morning post about how I solved the lunchbox problem.

The output sounds like the user wrote it. Because it learned the voice from the examples.

WHEN TO USE FEW-SHOT

Use it for:

Tone/voice matching (writing in your voice)
Format consistency (every output looks the same)
Structured outputs (lists, tables, briefs in specific format)
Niche tasks (where Claude doesn't have strong defaults)

Don't bother for:

Simple Q&A
Math
Tasks where Claude's default is already good

NUMBER OF EXAMPLES

1 example = "one-shot" — works but less reliable
2-3 examples = sweet spot for most tasks
5+ examples = diminishing returns + uses tokens

If 3 examples don't get the result, the issue isn't more examples — it's the examples you picked. Pick more representative ones.

ADVANCED: BAD + GOOD EXAMPLES

Here's a BAD example of what I DON'T want:
[bad example]

Here's a GOOD example of what I DO want:
[good example]

Now write one for: [your input]

Forcing the contrast sharpens the model's understanding. Especially useful when you've gotten 2-3 outputs that all miss in the same way.

3 THINGS YOU CAN'T SKIP

1. Use REAL examples. Don't fabricate examples to make Claude sound a certain way. Use stuff you've actually written or admired.

2. Make examples DIFFERENT enough. Three almost-identical examples don't teach the pattern as well as three varied ones.

3. Match input format to example format. If your examples are 3 sentences, ask for output in 3 sentences. If they're 200 words, ask for 200 words.

WHAT'S NEXT

Day 16 of 100. Tomorrow: Day 17 — Chain of Thought.

SAFETY CHECK

Same as Day 1.