Few-Shot Examples and Chain-of-Thought

Few-Shot Examples and Chain-of-Thought Prompting

Two of the most powerful prompt engineering techniques are few-shot prompting (showing examples) and chain-of-thought (asking Claude to reason step-by-step before answering).

Few-Shot Prompting

Provide input-output examples to demonstrate the exact format and behavior you want:

import anthropic

client = anthropic.Anthropic()

FEW_SHOT_SYSTEM = """
You extract structured data from unstructured customer feedback.
Always respond with valid JSON matching the schema shown in the examples.
"""

few_shot_messages = [
    {
        "role": "user",
        "content": "Feedback: 'The product arrived broken and customer support was rude.'"
    },
    {
        "role": "assistant",
        "content": '{"sentiment": "negative", "issues": ["damaged product", "poor support"], "priority": "high"}'
    },
    {
        "role": "user",
        "content": "Feedback: 'Delivery was fast and the packaging was excellent!'"
    },
    {
        "role": "assistant",
        "content": '{"sentiment": "positive", "issues": [], "priority": "low"}'
    },
    {
        "role": "user",
        "content": "Feedback: 'Good product but the setup instructions were confusing.'"
    },
]

response = client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=256,
    system=FEW_SHOT_SYSTEM,
    messages=few_shot_messages,
)

print(response.content[0].text)
# {"sentiment": "mixed", "issues": ["unclear instructions"], "priority": "medium"}

Chain-of-Thought Prompting

Asking Claude to reason before answering dramatically improves accuracy on complex tasks:

def solve_problem_with_cot(problem: str) -> dict:
    response = client.messages.create(
        model="claude-sonnet-4-5",
        max_tokens=2048,
        messages=[{
            "role": "user",
            "content": f"""Solve this problem step by step.

Problem: {problem}

First, think through the problem carefully in <thinking> tags.
Then provide your final answer in <answer> tags."""
        }],
    )

    text = response.content[0].text

    # Parse the structured response
    thinking = text.split("<thinking>")[1].split("</thinking>")[0].strip()
    answer = text.split("<answer>")[1].split("</answer>")[0].strip()

    return {"reasoning": thinking, "answer": answer}


result = solve_problem_with_cot(
    "A train leaves City A at 60 mph. Another leaves City B (300 miles away) "
    "at 40 mph heading toward City A. When do they meet?"
)

print("Reasoning:", result["reasoning"])
print("Answer:", result["answer"])

Extended Thinking (Claude 3.7+)

For the most demanding reasoning tasks, extended thinking gives Claude dedicated reasoning tokens:

response = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=16000,
    thinking={
        "type": "enabled",
        "budget_tokens": 10000,  # Tokens allocated for internal reasoning
    },
    messages=[{
        "role": "user",
        "content": "Prove that there are infinitely many prime numbers."
    }],
)

for block in response.content:
    if block.type == "thinking":
        print("Internal reasoning:", block.thinking[:200], "...")
    elif block.type == "text":
        print("Final answer:", block.text)

Few-shot examples anchor the output format; chain-of-thought anchors the reasoning quality.