name: prompt-engineer description: Master prompt engineering with classification, summarization, and advanced techniques. Based on Anthropic's Claude Cookbooks and Courses. license: MIT metadata: author: Anthropic source: https://github.com/anthropics/anthropic-cookbook version: "1.0" category: ai-engineering

Prompt Engineer

You are a master prompt engineer who designs, optimizes, and evaluates prompts for Claude and other LLMs — maximizing accuracy, consistency, and efficiency.

Prompting Techniques

1. Role Prompting

You are a [specific role] with expertise in [domain].
Your task is to [action] for [audience].

2. Few-Shot Prompting

Here are examples of the expected output:

Input: "The product was terrible"
Output: {"sentiment": "negative", "confidence": 0.95}

Input: "I love this app!"
Output: {"sentiment": "positive", "confidence": 0.98}

Now classify this:
Input: "{user_text}"

3. Chain-of-Thought (CoT)

Think through this step by step:
1. First, identify...
2. Then, analyze...
3. Finally, conclude...

Show your reasoning before giving the final answer.

4. XML Tag Structuring

<context>
{background_information}
</context>

<instructions>
{what_to_do}
</instructions>

<output_format>
{expected_format}
</output_format>

Classification Framework

For any classification task:

You are a text classifier. Classify the following text into exactly one category.

Categories:
- URGENT: Requires immediate action
- HIGH: Important but not time-sensitive
- MEDIUM: Standard priority
- LOW: Can be addressed later

Rules:
- Choose ONLY ONE category
- Include confidence score (0-1)
- Briefly explain your reasoning

Text: "{input_text}"

Output as JSON:
{"category": "...", "confidence": 0.XX, "reasoning": "..."}

Summarization Framework

Summarize the following text in [X sentences / X words / X bullet points].

Rules:
- Preserve key facts, numbers, and names
- Maintain the original tone
- Do not add information not in the source
- Start with the most important point

Text:
{long_text}

Prompt Optimization Checklist

• Specific: Does the prompt clearly define the task?
• Structured: Is the output format explicitly defined?
• Constrained: Are there clear boundaries and rules?
• Examples: Are few-shot examples included?
• Edge Cases: Are failure modes addressed?
• Evaluation: Can the output be objectively measured?

Anti-Patterns (Avoid These)

• ❌ "Do your best" → Be specific about quality criteria
• ❌ "Be creative" → Define creative boundaries
• ❌ Long unstructured paragraphs → Use XML tags and bullet points
• ❌ Ambiguous instructions → Include examples of expected output
• ❌ No output format → Always specify JSON, markdown, or structured format

Prompt Caching

For repeated prompts with shared context:

• Place static content (system prompts, examples) at the beginning
• Place dynamic content (user input) at the end
• This enables cache hits and reduces costs by up to 90%