Using Examples (Few-Shot Prompting)

import anthropic

client = anthropic.Anthropic()

# Few-shot via message history — the cleanest API pattern
response = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=256,
    system="You are a customer service agent for a software company. Classify support tickets into one of three categories: Bug / Feature Request / Account Issue. Respond with just the category name.",
    messages=[
        # Example 1
        {"role": "user", "content": "I can't log in — it says my password is wrong but I just reset it."},
        {"role": "assistant", "content": "Account Issue"},
        # Example 2
        {"role": "user", "content": "The export to CSV function crashes when I have more than 1000 rows."},
        {"role": "assistant", "content": "Bug"},
        # Example 3
        {"role": "user", "content": "It would be really helpful if I could schedule reports to send automatically."},
        {"role": "assistant", "content": "Feature Request"},
        # Actual input
        {"role": "user", "content": "The dark mode toggle doesn't save between sessions — every time I reload I have to switch it back."}
    ]
)

print(response.content[0].text)  # → "Bug"

Rewrite the following customer review response in our brand voice.
Brand voice: warm, direct, uses the customer's name, specific not generic,
ends with a forward-looking statement.

Here are two examples of our ideal responses:

<example>
<original>Thank you for your feedback. We're sorry for the inconvenience.</original>
<rewritten>Hi Sarah — we hear you, and that's genuinely frustrating. Our team is already looking at the sync issue you hit. You'll get an email the moment it's fixed. Thank you for taking the time to let us know.</rewritten>
</example>

<example>
<original>We appreciate your business and will look into this.</original>
<rewritten>James, thanks for flagging this — a 3-day delay is not okay and you deserve better. I've escalated this to our logistics team directly. Expect an update by tomorrow afternoon.</rewritten>
</example>

Now rewrite this response in the same voice:
<original>{original_response}</original>

TASK TYPE                    RECOMMENDED EXAMPLES  NOTES
─────────────────────────────────────────────────────────────────
Binary classification        2 (one per class)     One yes, one no
Multi-class (3-5 classes)    3-5 (one per class)   Cover all classes
Complex format matching      1-3                    One great example often sufficient
Style/tone matching          1-2                    Show don't tell
Extraction tasks             2-3                    Cover edge cases
Translation/transformation   2-3                    Include a tricky case
Creative writing style       3-5                    Show range of the style

Classify the sentiment of customer reviews as positive, negative, or neutral.

GOOD EXAMPLES:
Input: "Works exactly as described. Arrived on time."
Output: positive

Input: "Broke after one week. Terrible quality."
Output: negative

BAD EXAMPLES (do not do this):
Input: "It's fine I guess."
Output: positive  ← WRONG: "fine I guess" signals lukewarm disappointment, not satisfaction. This should be: neutral

Input: "Expensive but worth it."
Output: negative  ← WRONG: explicit statement of worth overrides the price complaint. This should be: positive

Classify this review: "Not what I expected, but I've gotten used to it."

from anthropic import Anthropic
import numpy as np

# Simplified example — in production use a vector DB (Pinecone, Weaviate, etc.)
def cosine_similarity(a, b):
    return np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b))

def select_relevant_examples(query: str, example_library: list, k: int = 3) -> list:
    """Select k most similar examples to query using embedding similarity."""
    client = Anthropic()
    # In practice: embed query and examples, find top-k by cosine similarity
    # This pseudocode illustrates the concept
    query_embedding = embed(query)
    scored = [(cosine_similarity(query_embedding, ex["embedding"]), ex)
              for ex in example_library]
    scored.sort(reverse=True)
    return [ex for _, ex in scored[:k]]

def classify_with_dynamic_examples(input_text: str) -> str:
    examples = select_relevant_examples(input_text, EXAMPLE_LIBRARY, k=3)
    messages = []
    for ex in examples:
        messages.append({"role": "user", "content": ex["input"]})
        messages.append({"role": "assistant", "content": ex["output"]})
    messages.append({"role": "user", "content": input_text})
    # ... make API call with dynamic messages