Advanced Prompting

<system_role>
You are a senior Python architect.
</system_role>

<context>
The user has a function that processes
large lists (100k+ items). It's slow.
</context>

<instructions>
Analyze the code and suggest optimizations.
Focus on algorithm complexity first, then
memory usage, then parallelization.
</instructions>

<rules>
- Don't suggest rewrites unless necessary
- Always explain the performance impact
- Cite specific lines of code
- Don't suggest external dependencies
</rules>

<output_format>
{
  "summary": "Brief overview",
  "optimizations": [
    {
      "issue": "What's slow",
      "fix": "How to fix it",
      "performance_gain": "Expected improvement",
      "code_location": "Line X-Y"
    }
  ]
}
</output_format>

<instructions>
You are a sentiment analysis system.
Analyze the text and respond with JSON
containing sentiment, confidence, and
supporting evidence.
</instructions>

<output_format>
{
  "sentiment": "positive|negative|neutral",
  "confidence": 0.0 to 1.0,
  "supporting_quote": "...",
  "explanation": "..."
}

IMPORTANT: Return ONLY valid JSON.
No markdown, no extra text. Just JSON.
</output_format>

Text to analyze: "I love this product!"

import json
import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=500,
    messages=[{"role": "user", "content": "..."}]
)

text = response.content[0].text

try:
    result = json.loads(text)
    print(f"Sentiment: {result['sentiment']}")
except json.JSONDecodeError as e:
    print(f"Failed to parse JSON: {e}")
    print(f"Raw response: {text}")

<instructions>
Summarize the meeting transcript using
this exact template. Fill in each section.
</instructions>

Meeting Summary Template:
──────────────────────────
Title: [Meeting name]
Date: [Date]
Duration: [Length in minutes]

Attendees: [List names]

Key Decisions:
1. [Decision 1]
2. [Decision 2]
3. [Decision 3]

Action Items:
- [ ] [Action 1] - Owner: [Name]
- [ ] [Action 2] - Owner: [Name]
- [ ] [Action 3] - Owner: [Name]

Next Steps: [What happens next]

Follow-up Meeting: [Date/Time if scheduled]
──────────────────────────

Now summarize this transcript: ...

import anthropic
import json

client = anthropic.Anthropic()

def chain_analysis(document):
    # Step 1: Extract key entities
    step1 = client.messages.create(
        model="claude-sonnet-4-5-20250929",
        max_tokens=500,
        messages=[{
            "role": "user",
            "content": f"Extract names, dates, and amounts from this document:\n{document}\n\nReturn JSON."
        }]
    )
    entities = json.loads(step1.content[0].text)

    # Step 2: Analyze sentiment of entities
    step2 = client.messages.create(
        model="claude-sonnet-4-5-20250929",
        max_tokens=500,
        messages=[{
            "role": "user",
            "content": f"Analyze sentiment of these names in context:\n{json.dumps(entities)}\n\nReturn JSON."
        }]
    )
    analysis = json.loads(step2.content[0].text)

    # Step 3: Generate summary
    step3 = client.messages.create(
        model="claude-sonnet-4-5-20250929",
        max_tokens=300,
        messages=[{
            "role": "user",
            "content": f"Summarize this analysis:\n{json.dumps(analysis)}\n\nKeep it under 100 words."
        }]
    )
    summary = step3.content[0].text

    return {
        "entities": entities,
        "analysis": analysis,
        "summary": summary
    }

result = chain_analysis("Alice sold 100 shares...")
print(json.dumps(result, indent=2))

<instructions>
Write a prompt that will make Claude classify
customer support tickets as urgent, normal,
or low-priority. The prompt should:
- Use XML tags for clarity
- Include 3 few-shot examples
- Define output format as JSON
- Include guardrails against bias
</instructions>

Return the complete prompt as a markdown code
block with proper formatting.

import anthropic

client = anthropic.Anthropic()

def generate_prompt(task_description):
    """Use Claude to write a prompt."""
    response = client.messages.create(
        model="claude-sonnet-4-5-20250929",
        max_tokens=1000,
        messages=[{
            "role": "user",
            "content": f"""Write a production-grade prompt for this task:

{task_description}

The prompt should use XML tags, include
few-shot examples, define output format,
and include guardrails. Return just the
prompt text, ready to use."""
        }]
    )
    return response.content[0].text

# Generate a prompt for data extraction
task = "Extract invoice data (vendor, amount, date) from unstructured text"
prompt = generate_prompt(task)
print(prompt)

# Now use the generated prompt
response = client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=500,
    messages=[{
        "role": "user",
        "content": prompt
    }]
)
print(response.content[0].text)