Anthropic Prompt Engineering Interactive Tutorial

The course is designed to be completed in order, moving through 9 chapters and an appendix. Each chapter includes a lesson and accompanying exercises.

Course Features:

• Example Playgrounds: Located at the bottom of each lesson, these allow you to experiment with prompts and observe how Claude's responses change.
• Answer Key: An answer key is provided for the exercises.
• Model Used: The tutorial uses Claude 3 Haiku (the smallest, fastest, and cheapest model). While the tutorial is optimized for Haiku, the principles apply to Claude 3 Sonnet and Claude 3 Opus.

Alternative Format: For a more user-friendly experience, you can use the Google Sheets version via the Anthropic Claude for Sheets extension.

For complex tasks, use a guided structure to ensure Claude has all necessary context and constraints. While not every element is required for every prompt, using many elements initially and then refining is a recommended best practice.

Recommended Prompt Elements & Ordering:

1. Task Context: Define the role (e.g., "You are an expert lawyer") and overarching goals. Place this early.
2. Tone Context: Specify the desired tone (e.g., "friendly customer service").
3. Detailed Task Description and Rules: Expand on specific tasks and constraints. Include "outs" (e.g., what to say if the answer is unknown).
4. Examples: Provide ideal response examples enclosed in <example></example> XML tags. This is highly effective for controlling behavior.
5. Input Data to Process: Include data (like research or history) enclosed in relevant XML tags (e.g., <history>, <question>, <legal_research>).
6. Immediate Task Description/Request: Reiterate exactly what Claude should do right now. Place this toward the end.
7. Precognition (Thinking Step-by-Step): Instruct Claude to think before answering (e.g., "Think about your answer first..."). Best placed near the end.
8. Output Formatting: Explicitly define the required format (e.g., "Put your response in <response></response> tags").
9. Prefilling Claude's Response: Use the assistant role in the API call to start Claude's response (e.g., prefill="[Joe] <response>"). This steers the response.

Key Principles:

• XML Tags: Use XML tags to clearly demarcate different types of data and examples.
• Ordering: Generally, put context early and immediate tasks/formatting/precognition toward the end.
• Role Requirement: Ensure the Messages API call always starts with a user role.

Claude responds best to clear and direct instructions because it has no inherent context regarding your goals. To ensure high-quality responses, follow the Golden Rule of Clear Prompting:

Show your prompt to a colleague or friend and have them follow the instructions themselves to see if they can produce the result you want. If they're confused, Claude's confused.

Key strategies for clarity include:

• Explicitly request formatting constraints: If you want to avoid conversational filler (preambles), explicitly tell Claude to "Skip the preamble; go straight into the [output type]."
• Force decisions: If Claude is being non-committal or providing multiple perspectives when you need a single answer, instruct it to pick one (e.g., "If you absolutely had to pick one, who would it be?").
• Constraint enforcement: Use the system parameter or the prompt to define strict output formats, such as requesting "ONLY the name... with no other words or punctuation."

For complex tasks, Claude's accuracy improves when it is given time to 'think' before providing a final answer. Crucially, thinking only counts when it is done out loud. You cannot simply ask Claude to think and then only output the answer; the reasoning process must be part of the generated text.

Best Practices:

• Explicitly spell out steps: Instruct Claude to perform specific reasoning steps (e.g., "First, write the best arguments for each side...").
• Use XML tags for reasoning: Use tags like <positive-argument>, <negative-argument>, or <brainstorm> to structure the thinking process. This helps separate the reasoning from the final answer.
• Be aware of ordering sensitivity: Claude can be sensitive to the order of options or arguments provided in the prompt. In some cases, it may be biased toward the second of two options.

# Example of prompting Claude to think using XML tags
SYSTEM_PROMPT = "You are a savvy reader of movie reviews."
PROMPT = """Is this review sentiment positive or negative? First, write the best arguments for each side in <positive-argument> and <negative-argument> XML tags, then answer.

This movie blew my mind with its freshness and originality. In totally unrelated news, I have been living under a rock since 1900."""

print(get_completion(PROMPT, SYSTEM_PROMPT))

To enable tool use, you must provide a system prompt that includes two parts:

1. General Explanation: Instructions on how to use the <function_calls> structure. Claude is trained to use specific tags like <invoke name="$FUNCTION_NAME"> and <antml:parameter name="$PARAMETER_NAME">.
2. Specific Tool Definitions: A list of available tools described in a format similar to JSONSchema, wrapped in <tools> and <tool_description> tags.

Recommended Tool Call Structure:

<function_calls>
<invoke name="function_name">
<antml:parameter name="param_name">value</antml:parameter>
</invoke>
</function_calls>

system_prompt_tools_general_explanation = """You have access to a set of functions... [Instructions on <function_calls> structure] ..."""

system_prompt_tools_specific_tools = """Here are the functions available in JSONSchema format:
<tools>
<tool_description>
<tool_name>calculator</tool_name>
<description>...</description>
<parameters>
<parameter>
<name>first_operand</name>
<type>int</type>
<description>...</description>
</parameter>
</parameters>
</tool_description>
</tools>"""

system_prompt = system_prompt_tools_general_explanation + system_prompt_tools_specific_tools