Claude Prompt Engineering

Claude is Anthropic's AI assistant designed to be helpful, harmless, and honest. It excels at long-context tasks, follows complex instructions precisely, and works best with well-structured prompts using XML-style tags.

When to Invoke This Skill

Use this skill when:

• Crafting prompts specifically for Claude/Anthropic models (default: Claude Sonnet 4.5)
• Working with long documents or large context (up to 1M tokens with Sonnet 4.5 beta)
• Using structured prompts with XML-style tags
• Implementing extended thinking for complex reasoning
• Requiring precise instruction following
• Building agentic workflows with parallel tool use

Claude's Identity & Characteristics

Attribute	Description
Personality	Helpful, harmless, honest
Constitutional AI	Built-in safety and ethical guidelines
Context Window	Up to 1M tokens (Sonnet 4.5 beta), 200K standard
Strengths	Long-context analysis, instruction following, document understanding, agentic tasks
Prompt Style	Structured, clear, XML-style formatting
Extended Thinking	Optional reasoning trace feature with tool use

Universal Prompting Techniques (Claude-Adapted)

1. Zero-Shot Prompting

Claude responds well to clear, direct zero-shot prompts.

Good Example:

Extract the key dates and events from the following text:

<text>
[paste text]
</text>

Output format: JSON with keys "date", "event", "description"

Less Effective:

Can you tell me what dates are in this text?

2. Few-Shot Prompting (Multishot)

Use well-formatted examples with XML structure.

<examples>
<example>
<input>
The conference is scheduled for March 15, 2025 in San Francisco.
</input>
<output>
{
  "date": "2025-03-15",
  "event": "conference",
  "location": "San Francisco"
}
</output>
</example>

<example>
<input>
Our next board meeting is on June 22nd.
</input>
<output>
{
  "date": "2025-06-22",
  "event": "board meeting"
}
</output>
</examples>

<input>
The product launches on September 1st in New York.
</input>

<output>

3. Chain-of-Thought Prompting

Claude has an Extended Thinking feature that shows reasoning (enabled via API, output in response):

I need to decide between these two job offers. Let me think through this step by step.

<job_offer_a>
[details]
</job_offer_a>

<job_offer_b>
[details]
</job_offer_b>

Please analyze both offers, show your reasoning, and provide a recommendation.

API enables extended thinking; response includes:

<thinking>
First, let me analyze the compensation...
Then consider the growth potential...
The work-life balance factors are...
The company stability differs by...
</thinking>

<answer>
[conclusion]
</answer>

4. Zero-Shot CoT

Simply add "Let's think step by step" or similar:

What's the most efficient route to visit all these cities?

Let's think step by step.

5. Prompt Chaining with XML Tags

Break complex tasks using XML delimiters:

Chain Step 1:

<task>
Extract relevant quotes from this document related to [topic].
</task>

<document>
[paste document]
</document>

<output_format>
<quotes>
<quote>[relevant quote 1]</quote>
<quote>[relevant quote 2]</quote>
</quotes>
</output_format>

Chain Step 2:

<task>
Summarize the extracted quotes and synthesize key insights.
</task>

<quotes>
[from previous response]
</quotes>

<output_format>
<summary>
[executive summary]
</summary>

<key_insights>
<insight>[insight 1]</insight>
<insight>[insight 2]</insight>
</key_insights>
</output_format>

6. ReAct Prompting

Use structured thought-action-observation cycles:

<question>
[research question]
</question>

<thought_1>
[what needs to be done first]
</thought_1>

<action_1>
[tool use or information gathering]
</action_1>

<observation_1>
[result from action]
</observation_1>

<thought_2>
[next step based on observation]
</thought_2>

<final_answer>
[conclusion]
</final_answer>

7. Tree of Thoughts

Use multiple reasoning paths with XML structure:

<problem>
[complex problem]
</problem>

<thought_paths>
<path_1>
<assumption>[approach 1]</assumption>
<reasoning>[step-by-step]</reasoning>
<conclusion>[result]</conclusion>
</path_1>

<path_2>
<assumption>[approach 2]</assumption>
<reasoning>[step-by-step]</reasoning>
<conclusion>[result]</conclusion>
</path_2>

<path_3>
<assumption>[approach 3]</assumption>
<reasoning>[step-by-step]</reasoning>
<conclusion>[result]</conclusion>
</path_3>
</thought_paths>

<synthesis>
[best path and why]
</synthesis>

Claude-Specific Best Practices

1. Use XML-Style Tags for Structure

Claude's official courses extensively use XML tags:

<context>
[background information]
</context>

<task>
[what needs to be done]
</task>

<examples>
[example inputs and outputs]
</examples>

<input>
[the actual input to process]
</input>

<output_format>
[expected format]
</output_format>

2. Structure Long Prompts Hierarchically

From Anthropic's official courses:

[TASK_CONTEXT]
Setting the stage and overall context

[TONE_CONTEXT]
How Claude should approach the task

[INPUT_DATA]
The actual data to work with

[EXAMPLES]
Few-shot examples

[TASK_DESCRIPTION]
Specific task details

[IMMEDIATE_TASK]
The immediate action to take

[OUTPUT_FORMATTING]
Expected output structure

3. Leverage Extended Thinking

For complex reasoning, enable Claude's extended thinking via the API:

API Syntax (Python SDK):

response = client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=4096,
    thinking={
        "type": "enabled",
        "budget_tokens": 8192
    },
    messages=[{
        "role": "user",
        "content": "Analyze this complex problem..."
    }]
)

Prompt-Side (XML structure for expected output):

<task>
[complex reasoning task]
</task>

<thinking>
[Claude will show its reasoning here]
</thinking>

<answer>
[final answer]
</answer>

Key Points:

• budget_tokens sets max tokens for reasoning (must be < max_tokens)
• Claude 4.5 returns summarized thinking by default
• First few lines are more verbose (useful for prompt engineering)
• You're billed for full thinking tokens, not summary tokens

4. Use System Prompts Effectively

System prompts set Claude's behavior:

System: You are a technical writer specializing in API documentation. Your responses are always:
- Clear and concise
- Technically accurate
- Formatted with Markdown
- Focused on developer needs

User: [your actual query]

5. Prefill Claude's Response

Guide the format by starting Claude's response:

<task>
Analyze this document and extract key findings.
</task>

<document>
[paste document]
</document>

<response>
<summary>
[Claude continues from here]

6. Cache Control for Long Prompts

Optimize for repeated prompts:

<cached_content cache_control="{\"type\":\"ephemeral\"}">
[large context that doesn't change]
</cached_content>

<task>
[specific task that varies]
</task>

7. Claude 4.5 Agent Features

Claude 4.5 introduces powerful agent capabilities:

Parallel Tool Use - Claude can use multiple tools simultaneously:

<task>
Analyze this data and create a visualization.
</task>

<tools>
- Web search for market data
- Code execution for analysis
- File write for chart output
</tools>

Claude will execute these in parallel when possible.

Memory Files - Claude can maintain knowledge across sessions:

<task>
When working on ongoing projects, create a memory file to track:
- Key decisions and rationale
- Project context and constraints
- Preferences and patterns
</task>

Claude will automatically update and reference memory files when given local file access.

Extended Thinking with Tools - Reasoning can pause to use tools:

# API: Enable extended thinking with tool use
response = client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=4096,
    thinking={
        "type": "enabled",
        "budget_tokens": 8192
    },
    tools=[web_search_tool],
    messages=[{
        "role": "user",
        "content": "Research [topic] and provide a comprehensive analysis."
    }]
)

# Claude can now use web search DURING extended thinking,
# alternating between reasoning and information gathering.

Anti-Patterns to Avoid

Anti-Pattern	Why It Fails	Better Approach
Ambiguous instructions	Claude follows literally	Be explicit about requirements
Missing output format	Unpredictable formatting	Always specify format
No structure in long prompts	Claude may lose track	Use XML tags and sections
Ignoring context window limits	Truncation issues	Be mindful of 200K/1M limits
Over-constraining creativity	Reduces Claude's helpfulness	Balance structure with flexibility

Quick Reference Templates

Document Analysis

<task>
[specific analysis task]
</task>

<document>
[paste document]
</document>

<output_format>
[expected structure]
</output_format>

Code Generation with Examples

<task>
Write a function that [description]
</task>

<requirements>
[specific requirements]
</requirements>

<examples>
<example>
<input>[input example]</input>
<output>[expected output]</output>
</example>
</examples>

<output_format>
[code in specified language]
</output_format>

Data Extraction

<task>
Extract [specific fields] from the following text
</task>

<input_text>
[paste text]
</input_text>

<output_format>
JSON with keys: [list keys]
</output_format>

Extended Thinking

# API: Enable extended thinking
response = client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=4096,
    thinking={
        "type": "enabled",
        "budget_tokens": 8192
    },
    messages=[{
        "role": "user",
        "content": """
<task>
[complex reasoning task]
</task>

Please show your reasoning step by step, then provide the final answer.
"""
    }]
)

# Response will include <thinking> block followed by <answer>

Model Capabilities Reference

Feature	Claude Sonnet 4.5	Claude Haiku 4.5	Claude Opus 4.5	Notes
Context Window	200K / 1M (beta)	200K	200K	Sonnet 4.5: 1M with beta header
Extended Thinking	✅ Yes	✅ Yes	✅ Yes	With tool use support
Max Output	64K tokens	64K tokens	64K tokens	Unified across 4.5
Vision	✅ Yes	✅ Yes	✅ Yes	Image analysis
Parallel Tool Use	✅ Yes	✅ Yes	✅ Yes	Claude 4.5 feature
Memory Files	✅ Yes	✅ Yes	✅ Best	Local file knowledge
Code	✅ Excellent	✅ Good	✅ Best	Opus 4.5: SOTA coding
Speed	Fast	Fastest	Moderate	Default: Sonnet 4.5

Recommendation: Start with Claude Sonnet 4.5 - best balance of intelligence, speed, and cost for most use cases. Use Opus 4.5 for complex coding, Haiku 4.5 for speed-critical tasks.

Migration Notes (Claude 3 → 4.5)

If you're migrating from Claude 3.x to Claude 4.5:

Change	Impact	Action
Default model	Sonnet 3.5 → Sonnet 4.5	Update model IDs in code
Context window	200K → 1M (beta) available	Requires beta header for 1M
Parallel tools	New capability	Update prompts to leverage parallel execution
Memory files	New capability	Grant file access for persistent knowledge
Extended thinking + tools	New capability	Can now use tools during reasoning
Max output	8K → 64K tokens	Adjust output expectations

API Migration:

# Old (Claude 3.5)
model="claude-sonnet-3-5-20240620"

# New (Claude 4.5)
model="claude-sonnet-4-5-20250929"  # or use alias "claude-sonnet-4-5"

Most prompting techniques remain unchanged—XML tags, system prompts, and structured outputs work identically.

Prompt Element Checklist

When creating Claude prompts, consider including:

• Task Context: Overall purpose and setting
• Tone Context: How Claude should approach it
• Input Data: The actual content to process
• Examples: Few-shot demonstrations (if needed)
• Task Description: Specific instructions
• Immediate Task: What to do right now
• Output Format: Expected structure
• Prefill: Start of Claude's response (optional)

See Also

• references/basics.md - Foundational Claude prompting concepts
• references/techniques.md - Detailed technique explanations
• references/xml-formatting.md - XML tag patterns and usage
• references/patterns.md - Reusable Claude prompt patterns
• references/examples.md - Concrete examples from Anthropic courses
• grok-prompting skill - For Grok/xAI-specific guidance
• gemini-prompting skill - For Google Gemini-specific guidance