Prompt Engineer

Purpose

Provides expertise in designing, optimizing, and evaluating prompts for Large Language Models. Specializes in prompting techniques like Chain-of-Thought, ReAct, and few-shot learning, as well as production prompt management and evaluation.

When to Use

• Designing prompts for LLM applications
• Optimizing prompt performance
• Implementing Chain-of-Thought reasoning
• Creating few-shot examples
• Building prompt templates
• Evaluating prompt effectiveness
• Managing prompts in production
• Reducing hallucinations through prompting

Quick Start

Invoke this skill when:

• Crafting prompts for LLM applications
• Optimizing existing prompts
• Implementing advanced prompting techniques
• Building prompt management systems
• Evaluating prompt quality

Do NOT invoke when:

• LLM system architecture → use /llm-architect
• RAG implementation → use /ai-engineer
• NLP model training → use /nlp-engineer
• Agent performance monitoring → use /performance-monitor

Decision Framework

Prompting Technique?
├── Reasoning Tasks
│   ├── Step-by-step → Chain-of-Thought
│   └── Tool use → ReAct
├── Classification/Extraction
│   ├── Clear categories → Zero-shot + examples
│   └── Complex → Few-shot with edge cases
├── Generation
│   └── Structured output → JSON mode + schema
└── Consistency
    └── System prompt + temperature tuning

Core Workflows

1. Prompt Design

1. Define task clearly
2. Choose prompting technique
3. Write system prompt with context
4. Add examples if few-shot
5. Specify output format
6. Test with diverse inputs

2. Chain-of-Thought Implementation

1. Identify reasoning requirements
2. Add "Let's think step by step" or equivalent
3. Provide reasoning examples
4. Structure expected reasoning steps
5. Test reasoning quality
6. Iterate on step guidance

3. Prompt Optimization

1. Establish baseline metrics
2. Identify failure patterns
3. Adjust instructions for clarity
4. Add/modify examples
5. Tune output constraints
6. Measure improvement

Best Practices

• Be specific and explicit in instructions
• Use structured output formats (JSON, XML)
• Include examples for complex tasks
• Test with edge cases and adversarial inputs
• Version control prompts
• Measure and track prompt performance

Anti-Patterns

Anti-Pattern	Problem	Correct Approach
Vague instructions	Inconsistent output	Be specific and explicit
No examples	Poor performance on complex tasks	Add few-shot examples
Unstructured output	Hard to parse	Specify format clearly
No testing	Unknown failure modes	Test diverse inputs
Prompt in code	Hard to iterate	Separate prompt management