Prompt Engineering with Claude: Tutorial Summary

System: "You are a cat."
User: "What do you think about skateboarding?"

System: "You are a logic bot designed to answer complex logic problems."

System: "You are a savvy reader of movie reviews."

# Variable content
SENTENCES = """- I like how cows sound
- This sentence is about spiders
- This sentence is about pigs"""

# WITHOUT XML tags - confusing
PROMPT = f"""Below is a list. Tell me the second item.
- Each is about an animal, like rabbits.
{SENTENCES}"""
# Claude might think "rabbits" is part of the list!

# WITH XML tags - clear
PROMPT = f"""Below is a list. Tell me the second item.
- Each is about an animal, like rabbits.
<sentences>
{SENTENCES}
</sentences>"""
# Now Claude knows the sentences variable is the data, not the instruction

"Write a haiku about cats. Put it in <haiku> tags."

Claude responds:
<haiku>
Soft paws tread softly
Whiskers twitch in morning light
Purring fills the room
</haiku>

messages=[
  {"role": "user", "content": "Write a haiku about cats. Put it in <haiku> tags."},
  {"role": "assistant", "content": "<haiku>"}
]

Claude continues directly:
<haiku>
Soft paws tread softly...

PROMPT = "Write a haiku using JSON with keys 'first_line', 'second_line', 'third_line'."
PREFILL = "{"

Claude responds with valid JSON:
{
  "first_line": "Soft paws tread softly",
  "second_line": "Whiskers twitch in morning light",
  "third_line": "Purring fills the room"
}

EMAIL = "Hi Zack, ping on that prompt."
ADJECTIVE = "olde english"

# Use f-string placeholders for variables AND XML tags for structure
PROMPT = f"Make this email more {ADJECTIVE}: <email>{EMAIL}</email>. Write the result in <{ADJECTIVE}_email> tags."

# Prefill to ensure format
PREFILL = f"<{ADJECTIVE}_email>"

message = client.messages.create(
    model=MODEL_NAME,
    max_tokens=2000,
    stop_sequences=["</haiku>"],  # Stop after closing tag
    messages=[...]
)

Q: "Name a famous movie starring an actor born in 1956."
A: "The Shawshank Redemption starring Tim Robbins" [INCORRECT - wrong birth year]

Q: "Name a famous movie starring an actor born in 1956. First brainstorm actors
and birth years in <brainstorm> tags."

A: "<brainstorm>
- Tom Hanks was born in 1956
- He starred in Forrest Gump, Saving Private Ryan, Cast Away
</brainstorm>

Forrest Gump starring Tom Hanks (born 1956)" [CORRECT]

Q: "Will Santa bring me presents on Christmas?"
A: "Santa Claus is a folkloric figure. The tradition involves..."

"Please complete the conversation by writing the next line, speaking as 'A'.

Q: Is the tooth fairy real?
A: Of course, sweetie. Wrap up your tooth and put it under your pillow tonight.
   There might be something waiting for you in the morning.

Q: Will Santa bring me presents on Christmas?"

A: Absolutely! Make sure to leave out some cookies and milk for Santa on
   Christmas Eve, and he'll bring presents for good boys and girls like you!

Text: "Dr. Liam Patel is a neurosurgeon. Olivia Chen is an architect."
<individuals>
1. Dr. Liam Patel [NEUROSURGEON]
2. Olivia Chen [ARCHITECT]
</individuals>

Text: "Chef Oliver Hamilton runs Green Plate. Elizabeth Chen is a librarian."
<individuals>
1. Oliver Hamilton [CHEF]
2. Elizabeth Chen [LIBRARIAN]
</individuals>

Now extract from this text: "Laura Simmons is a farmer. Kevin Alvarez teaches dance."

# Few-shot examples + XML tags + prefilling
PROMPT = """Classify emails into categories.

Example:
Email: <email>My product is broken.</email>
Category: (B) Broken or defective item

Now classify:
Email: <email>{user_email}</email>
Category:"""

PREFILL = "("  # Start the category format

System: "You are a careful analyst who only makes claims supported by provided sources."

User: "Here is a research paper:
<paper>
[Paper content]
</paper>

Summarize the key findings. For each finding, cite the specific section of
the paper. If something is unclear or not stated in the paper, explicitly
say so."

System: "You are an email classification expert."

User: "Classify emails into these categories:
(A) Pre-sale question
(B) Broken or defective item
(C) Billing question
(D) Other

Examples:
<email>My Mixmaster4000 is making strange noises and smells like burning plastic.</email>
<thinking>This describes a defective product needing replacement.</thinking>
<answer>B</answer>

<email>Can I use the Mixmaster4000 to mix paint?</email>
<thinking>This is asking about product capabilities before purchase.</thinking>
<answer>A</answer>

Now classify:
<email>{user_email_text}</email>"

Assistant (prefill): "<thinking>"

User: "Extract names and occupations from text.

<text>Dr. Lisa Wong leads the research team. Marcus Johnson manages operations.</text>
<individuals>
1. Dr. Lisa Wong [RESEARCHER]
2. Marcus Johnson [OPERATIONS MANAGER]
</individuals>

<text>{user_text}</text>"

Assistant (prefill): "<individuals>"

System: "You are an expert business communications writer."

User: "Rewrite emails to be more professional.

Example:
<original>yo can u send me that file asap thx</original>
<professional>Good morning,

Could you please send me the file at your earliest convenience? Thank you for your assistance.

Best regards</professional>

Now rewrite:
<original>{user_email}</original>"

Assistant (prefill): "<professional>"

System: "You are a senior financial analyst with expertise in earnings reports."

User: "Analyze this earnings report:
<report>{earnings_data}</report>

Process:
1. Identify key metrics in <metrics> tags
2. Evaluate trends and concerns in <analysis> tags
3. Provide recommendations in <recommendations> tags

Base all analysis on data from the report. Cite specific figures."

Assistant (prefill): "<metrics>"

BAD:  "Make this email better"

GOOD: "Make this email more professional by:
1. Removing casual language and slang
2. Adding formal greetings and closings
3. Organizing into clear paragraphs
4. Using complete sentences"

# BAD - No clear boundaries
EMAIL = "Send me the report now!"
PROMPT = f"Rewrite this email nicely: {EMAIL} but keep it short"
# Claude might think "but keep it short" is part of the email

# GOOD - Clear boundaries
EMAIL = "Send me the report now!"
PROMPT = f"Rewrite this email nicely: <email>{EMAIL}</email> Keep the rewrite short."

BAD:  "Think carefully about this math problem and give the answer"

GOOD: "First show your step-by-step work in <work> tags, then give the final
       answer in <answer> tags"

# BAD - Inconsistent format
Example 1: "Input: X, Output: Y"
Example 2: "Q: X\nA: Y"
Example 3: "X => Y"

# GOOD - Consistent format
Example 1: <input>X</input><output>Y</output>
Example 2: <input>X</input><output>Y</output>
Example 3: <input>X</input><output>Y</output>

BETTER: "Is this review positive or negative? Consider both possibilities
         equally before deciding. Show your reasoning for each side first."

# SUBOPTIMAL
User: "You are an expert analyst. Please analyze this data carefully and professionally..."

# BETTER
System: "You are an expert data analyst. Always show your reasoning and cite specific data points."
User: "Analyze this data: <data>{data}</data>"

# BAD - Too much at once
"Read this document, extract key points, summarize them, translate to Spanish,
format as a presentation, and create discussion questions"

# GOOD - Sequential steps
Prompt 1: "Extract key points from: <document>{doc}</document>"
Prompt 2: "Summarize these points: <points>{extracted_points}</points>"
Prompt 3: "Translate to Spanish: <summary>{summary}</summary>"
... etc

import anthropic

client = anthropic.Anthropic(api_key=API_KEY)

message = client.messages.create(
    model="claude-3-haiku-20240307",
    max_tokens=2000,
    temperature=0.0,
    system="Your system prompt here",
    messages=[
        {"role": "user", "content": "Your prompt here"},
        {"role": "assistant", "content": "Optional prefill text"}
    ],
    stop_sequences=["</tag>"]  # Optional
)

response = message.content[0].text

def get_completion(prompt: str, system_prompt="", prefill=""):
    message = client.messages.create(
        model="claude-3-haiku-20240307",
        max_tokens=2000,
        temperature=0.0,
        system=system_prompt,
        messages=[
            {"role": "user", "content": prompt},
            {"role": "assistant", "content": prefill}
        ]
    )
    return message.content[0].text

# Usage
response = get_completion(
    prompt="What is 2+2?",
    system_prompt="You are a math tutor.",
    prefill="Let me calculate:"
)