Appendix 02: Anthropic Claude Prompting Guide Analysis

tech
prompt-engineering
anthropic
claude
extended-thinking
Deep analysis of Anthropic’s official prompting guides for Claude Sonnet 4, Opus 4.6, and Extended Thinking models with extracted techniques and examples
Author

Dario Airoldi

Published

January 20, 2026

Appendix 02: Anthropic Claude Prompting Guide Analysis

This appendix provides a comprehensive analysis of Anthropic’s official prompting documentation, extracting key techniques, patterns, and recommendations for Claude Sonnet 4, Claude Opus 4.6, and Extended Thinking models.

Guide Version: This analysis is based on Anthropic documentation as of 2026-01-20. Official guides may have been updated since this analysis. Always verify with the official documentation.

Table of Contents

📊 Model Overview

Claude Model Family

Model Best For Context Window Key Characteristic
Claude Sonnet 4 Balanced tasks, coding, analysis 200K tokens Fast, capable, cost-effective
Claude Opus 4 Complex reasoning, nuanced writing 200K tokens Highest capability
Claude Opus 4.6 Frontier agentic tasks 200K tokens Enhanced multi-step reasoning, tool use, agentic reliability
Claude with Extended Thinking STEM, complex problems 200K tokens Internal reasoning before response

Core Philosophy

Anthropic’s prompting guidance centers on one key insight:

“Think of Claude as a brilliant but very new employee (with amnesia) who needs explicit instructions.”

Claude doesn’t have context on your norms, styles, guidelines, or preferred ways of working. The more precisely you explain what you want, the better Claude’s response will be.

Prerequisites Before Prompting

Before starting prompt engineering, Anthropic recommends:

  1. ✅ Clear success criteria defined for your use case
  2. ✅ Empirical testing methods established
  3. ✅ First draft prompt ready to improve

🎯 The Golden Rule of Clarity

Show your prompt to a colleague, ideally someone who has minimal context on the task, and ask them to follow the instructions. If they’re confused, Claude will likely be too.

How to Be Clear, Contextual, and Specific

1. Provide Contextual Information

Give Claude the context it needs to understand the task:

Context Type Example
Purpose “This output will be used for a customer-facing FAQ”
Audience “The reader is a senior developer familiar with TypeScript”
Workflow position “This is step 2 of a 4-step review process”
Success criteria “A successful response includes code examples and error handling”

2. Be Specific About Requirements

❌ Vague: “Analyze this code”

✅ Specific: “Analyze this code for:

  • Security vulnerabilities (especially SQL injection)
  • Performance bottlenecks
  • Violations of our coding standards Output a numbered list with severity ratings.”

3. Use Sequential Steps

Provide instructions as numbered steps:

1. Read the customer complaint
2. Identify the main issue category (billing, technical, shipping)
3. Extract the customer's emotional tone
4. Draft a response that addresses the issue
5. Include a resolution offer if appropriate

🏷️ XML Tags for Structure

Claude responds exceptionally well to XML tags for organizing prompts and defining clear boundaries.

Why XML Works Well

  • Creates unambiguous section boundaries
  • Prevents prompt injection from user content
  • Makes parsing of outputs straightforward
  • Allows metadata via attributes

Basic Structure

<role>
You are a technical documentation specialist reviewing API docs.
</role>

<context>
The API serves a financial services application with strict compliance requirements.
</context>

<task>
Review the following API documentation for completeness.
</task>

<document>
{{user_provided_document}}
</document>

<output_format>
Return a JSON object with:
- completeness_score (0-100)
- missing_sections (array)
- recommendations (array)
</output_format>

Common Tag Patterns

Tag Purpose Example Use
<role> Define persona/expertise <role>You are a senior security auditor.</role>
<context> Background information <context>This is a healthcare application.</context>
<task> Specific instruction <task>Identify HIPAA compliance issues.</task>
<document> User-provided content <document>{{file_content}}</document>
<examples> Few-shot examples Wrap input/output pairs
<thinking> Request reasoning output <thinking>[Claude's reasoning]</thinking>
<answer> Final response section Separate from thinking
<constraints> Rules and limitations <constraints>Max 500 words.</constraints>

Nested Structure Example

<task>
Analyze customer feedback and generate improvement recommendations.
</task>

<feedback_data>
  <feedback id="1" date="2026-01-15">
    Great product, but shipping was slow.
  </feedback>
  <feedback id="2" date="2026-01-16">
    Love the quality! Would buy again.
  </feedback>
  <feedback id="3" date="2026-01-17">
    Customer service was unhelpful with my return.
  </feedback>
</feedback_data>

<output_format>
<analysis>
  <theme name="[theme]">
    <frequency>[count]</frequency>
    <sentiment>[positive/negative/neutral]</sentiment>
    <recommendation>[action item]</recommendation>
  </theme>
</analysis>
</output_format>

💭 Chain of Thought Prompting

For complex tasks, letting Claude think step-by-step dramatically improves performance.

Benefits of CoT

Benefit Description
Accuracy Reduces errors in math, logic, analysis
Coherence Leads to more organized responses
Debugging Reveals where prompts may be unclear

When to Use (and Not Use)

✅ Use CoT for: Tasks a human would need to think through—complex math, multi-step analysis, writing complex documents, decisions with many factors.

❌ Avoid CoT for: Simple tasks where it adds unnecessary latency.

Three Levels of CoT

Level 1: Basic (Minimal Guidance)

Think step-by-step about this problem:
{{problem}}

Limitation: Lacks guidance on HOW to think.

Level 2: Guided (Structured Steps)

Analyze this code for security vulnerabilities.

Think through:
1. Input validation issues
2. Authentication weaknesses  
3. Data exposure risks
4. Injection vulnerabilities

Then provide your findings.

Benefit: Directs Claude’s reasoning process.

Level 3: Structured (Separated Output)

Analyze this financial report for investment risks.

<thinking>
[Your step-by-step analysis here]
</thinking>

<answer>
[Your final recommendation here]
</answer>

Benefit: Easy to parse and separate reasoning from conclusions.

Complete CoT Example

<task>
Draft a personalized donor thank-you email.
</task>

<donor_info>
Name: Sarah Chen
Donation: $500
History: 3rd year donor, total $1,200
Cause: Education programs
</donor_info>

<instructions>
Think through:
1. What makes this donor special?
2. What impact did their donation have?
3. How to personalize the message?
4. What's an appropriate call-to-action?

Then write the email.
</instructions>

<thinking>
[Claude's reasoning process]
</thinking>

<email>
[Final personalized email]
</email>

🚀 Opus 4.6 Optimization Strategies

Claude Opus 4.6 is Anthropic’s frontier model, designed for the most demanding agentic and analytical tasks. It builds on Opus 4’s strengths with significant improvements in multi-step tool use, instruction fidelity, and long-context recall.

When to choose Opus 4.6

Scenario Why Opus 4.6
Multi-step agentic workflows Excels at sequencing tool calls correctly across many steps
Dense system prompts Supports highly detailed instructions with minimal drift
Extended context fidelity Better recall across long contexts compared to Sonnet 4
Complex analysis + writing Produces more nuanced, thorough outputs than Opus 4

Prompting tips for Opus 4.6

  • Don’t simplify instructions — Opus 4.6 handles dense, layered prompts well. Where Sonnet benefits from concise guidance, Opus 4.6 thrives with comprehensive detail.
  • Trust tool sequencing — In agentic workflows, Opus 4.6 reliably chains tool calls in the correct order. Provide the goal and constraints rather than scripting each step.
  • Leverage long context — Place reference material early in the context window. Opus 4.6 maintains better recall across the full 200K token window than prior Opus models.
  • Pair with Extended Thinking — For maximum reasoning depth on complex analysis tasks, enable Extended Thinking mode with Opus 4.6. The combination delivers the strongest results for STEM, planning, and constraint satisfaction problems.

⚠️ Note: Verify these details against the official Anthropic Opus 4.6 release documentation before publishing. Model capabilities may evolve.

🧠 Extended Thinking Mode

Extended thinking is Claude’s internal reasoning capability for complex problems, similar to OpenAI’s reasoning models.

Key Differences from Standard CoT

Aspect Standard CoT Extended Thinking
Visibility Shown in output Internal (summary optional)
Prompting Must request explicitly Enabled via API parameter
Instruction style Step-by-step guidance High-level goals preferred
Budget N/A Token budget configurable

Technical Considerations

  • Minimum budget: 1024 tokens
  • For >32K thinking: Use batch processing
  • Language: Best in English (outputs can be any language)
  • Below minimum: Use standard CoT with <thinking> tags instead

Prompting for Extended Thinking

✅ DO: High-Level Instructions

Please think about this math problem thoroughly and in great detail.
Consider multiple approaches and show your complete reasoning.
Try different methods if your first approach doesn't work.

❌ DON’T: Over-Prescribe Steps

Think through this math problem step by step:
1. First, identify the variables
2. Then, set up the equation
3. Next, solve for x
...

Why? Claude’s creativity in approaching problems may exceed your ability to prescribe the optimal thinking process.

When to Increase Thinking Budget

Extended thinking excels at:

  • Complex STEM problems - Mathematical proofs, physics
  • Constraint optimization - Scheduling, resource allocation
  • Thinking frameworks - Multi-perspective analysis
  • Code with edge cases - Algorithms requiring verification

Multishot with Extended Thinking

You can provide examples showing how to think:

I'll show you how to solve a math problem, then solve a similar one.

Problem 1: What is 15% of 80?

<thinking>
To find 15% of 80:
1. Convert 15% to a decimal: 15% = 0.15
2. Multiply: 0.15 × 80 = 12
</thinking>

The answer is 12.

Now solve this one:
Problem 2: What is 35% of 240?

Self-Verification

Ask Claude to check its work:

Write a function to calculate the factorial of a number.
Before you finish, please verify your solution with test cases for:
- n=0
- n=1
- n=5
- n=10
And fix any issues you find.

What NOT to Do

❌ Don’t pass thinking back: Don’t include Claude’s extended thinking in subsequent prompts—this doesn’t improve performance.

❌ Don’t prefill: Prefilling assistant responses with extended thinking is explicitly not allowed.

❌ Don’t manually edit: Changing model output after thinking can degrade results.

📚 Multishot Prompting

Providing examples (multishot/few-shot) is one of the most effective ways to guide Claude’s output.

When Examples Help Most

  • Establishing specific output formats
  • Teaching domain-specific patterns
  • Demonstrating edge case handling
  • Calibrating response length and style

Example Structure

<examples>
  <example id="1">
    <input>
      Customer complaint: "My order arrived damaged"
    </input>
    <output>
      Category: Shipping
      Sentiment: Negative
      Priority: High
      Suggested action: Offer replacement + discount
    </output>
  </example>
  
  <example id="2">
    <input>
      Customer feedback: "Love the new features!"
    </input>
    <output>
      Category: Product
      Sentiment: Positive
      Priority: Low
      Suggested action: Thank and request review
    </output>
  </example>
</examples>

<task>
Analyze this customer message:
<input>
{{customer_message}}
</input>
</task>

Best Practices

  1. Use 2-5 diverse examples covering different scenarios
  2. Maintain consistent formatting across all examples
  3. Include edge cases if relevant
  4. Use IDs to clearly link inputs with outputs

👤 System Prompts and Roles

System prompts establish Claude’s persona and high-level behavior.

Effective Role Definition

<system>
You are a senior security consultant specializing in web application security.
You have 15 years of experience in penetration testing and code review.
Your communication style is direct and actionable.
You always cite specific line numbers when identifying issues.
</system>

Role Components

Component Purpose Example
Expertise Domain knowledge “specializing in financial compliance”
Experience Calibrate depth “15 years of experience”
Style Communication tone “concise and technical”
Behaviors Specific actions “always provide examples”

📝 Long Context Tips

Claude excels at processing long documents (up to 200K tokens), but benefits from guidance.

Placement Strategies

  1. Critical instructions FIRST: Place essential guidance at prompt start
  2. Large documents MIDDLE: Reference material in the body
  3. Specific queries LAST: Ask questions after providing context

Document Delineation

<documents>
  <document id="1" title="Q1 Financial Report">
    {{document_1_content}}
  </document>
  
  <document id="2" title="Q2 Financial Report">
    {{document_2_content}}
  </document>
</documents>

<task>
Compare the revenue trends between Q1 and Q2, citing specific figures from each document.
</task>

Needle-in-Haystack Queries

For finding specific information in large contexts:

The following documents contain our company's HR policies.

<documents>
{{large_policy_documents}}
</documents>

Find and quote the exact policy text regarding:
1. Remote work eligibility requirements
2. Equipment reimbursement limits

Cite the document name and section for each answer.

✂️ Prefilling Claude’s Response

You can start Claude’s response to control format and reduce unwanted preamble.

Basic Prefilling

response = client.messages.create(
    model="claude-sonnet-4-20260514",
    messages=[
        {"role": "user", "content": "List the top 3 programming languages for data science."},
        {"role": "assistant", "content": "1."}  # Prefill starts the list
    ]
)

JSON Output Prefilling

response = client.messages.create(
    model="claude-sonnet-4-20260514",
    messages=[
        {"role": "user", "content": "Extract the person's name and age from: 'John is 32 years old.'"},
        {"role": "assistant", "content": "{"}  # Forces JSON output
    ]
)

XML Output Prefilling

response = client.messages.create(
    model="claude-sonnet-4-20260514",
    messages=[
        {"role": "user", "content": "Analyze the sentiment of this review: 'Great product!'"},
        {"role": "assistant", "content": "<analysis>"}  # Forces structured output
    ]
)

Note: Prefilling is NOT allowed when extended thinking is enabled.

🔗 Chain Complex Prompts

For multi-step workflows, break prompts into a chain where each step feeds the next.

Chaining Pattern

Prompt 1: Extract key information
    ↓
Output 1 → Input 2
    ↓
Prompt 2: Analyze extracted information
    ↓
Output 2 → Input 3
    ↓
Prompt 3: Generate final recommendation

Example: Document Analysis Chain

Step 1: Extraction

<task>Extract all financial figures from this document.</task>
<document>{{report}}</document>
<output_format>JSON with figure name, value, and context.</output_format>

Step 2: Analysis

<task>Identify trends and anomalies in these figures.</task>
<data>{{step_1_output}}</data>
<output_format>List of observations with significance ratings.</output_format>

Step 3: Recommendation

<task>Based on this analysis, provide investment recommendations.</task>
<analysis>{{step_2_output}}</analysis>
<output_format>Executive summary with action items.</output_format>

🔧 Practical Examples

Example 1: Customer Feedback Anonymization

<role>
You are a data privacy specialist responsible for anonymizing customer feedback
before it's shared with product teams.
</role>

<task>
Anonymize the following customer feedback by:
1. Replacing names with [CUSTOMER]
2. Removing specific dates (replace with [DATE])
3. Generalizing location information
4. Preserving the sentiment and actionable insights
</task>

<feedback>
Hi, I'm John Smith from Seattle. On January 15th, I ordered a laptop 
but received it damaged. I called Sarah in customer service and she 
was very helpful. Please ship a replacement to my address at 123 Main St.
</feedback>

<output_format>
Return only the anonymized text with no additional commentary.
</output_format>

Example 2: Marketing Email with CoT

<task>
Create a marketing email for our new productivity app targeting remote workers.
</task>

<product_info>
- App name: FocusFlow
- Key feature: AI-powered distraction blocking
- Price: $9.99/month
- Launch date: February 2026
</product_info>

<thinking>
Consider:
1. Pain points of remote workers (distractions, work-life balance)
2. How FocusFlow addresses these
3. Compelling subject line options
4. Call-to-action that drives sign-ups
</thinking>

<email>
Subject: [subject line]

[email body]
</email>

Example 3: Incident Response Analysis

<role>
You are a senior DevOps engineer investigating a production incident.
</role>

<context>
Our e-commerce platform experienced a 45-minute outage affecting checkout.
You need to analyze the logs and provide a root cause analysis.
</context>

<logs>
{{incident_logs}}
</logs>

<task>
Analyze the logs and provide:
1. Timeline of events leading to the outage
2. Root cause identification
3. Contributing factors
4. Immediate remediation steps taken
5. Long-term prevention recommendations
</task>

<output_format>
Use headers for each section. Include specific log timestamps as evidence.
</output_format>

⚠️ Common Pitfalls

Pitfall 1: Vague Instructions

❌ Wrong:

Help me with this code.

✅ Correct:

Review this Python function for:
1. Potential bugs
2. Performance issues
3. Violations of PEP 8

Suggest specific fixes for each issue found.

Pitfall 2: Missing Context

❌ Wrong:

Write a response to this email.

✅ Correct:

<context>
You are the customer success manager at a SaaS company.
The customer has been with us for 2 years and is considering churning.
</context>

<task>
Write a personalized response addressing their concerns and offering solutions.
</task>

<email>
{{customer_email}}
</email>

Pitfall 3: Inconsistent Example Formatting

❌ Wrong:

Example 1: Input "hello" gives greeting
Example 2:
  input: goodbye
  result: farewell

✅ Correct:

<example id="1">
  <input>hello</input>
  <output>greeting</output>
</example>
<example id="2">
  <input>goodbye</input>
  <output>farewell</output>
</example>

Pitfall 4: Over-Prescribing Extended Thinking

❌ Wrong (with extended thinking enabled):

Think step by step:
1. First consider X
2. Then analyze Y
3. Finally conclude Z

✅ Correct:

Think deeply about this problem. Consider multiple approaches
and verify your solution before answering.

Pitfall 5: Passing Thinking Back

❌ Wrong:

# Don't do this
second_response = client.messages.create(
    messages=[
        {"role": "user", "content": f"Based on your thinking: {first_response.thinking}..."}
    ]
)

✅ Correct:

# Use only the final output
second_response = client.messages.create(
    messages=[
        {"role": "user", "content": f"Based on your analysis: {first_response.content}..."}
    ]
)

📚 References

Official Documentation

Interactive Resources

Parent Article: 08.00 How to Optimize Prompts for Specific Models