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

• Define task clearly

• Choose prompting technique

• Write system prompt with context

• Add examples if few-shot

• Specify output format

• Test with diverse inputs

2. Chain-of-Thought Implementation

• Identify reasoning requirements

• Add "Let's think step by step" or equivalent

• Provide reasoning examples

• Structure expected reasoning steps

• Test reasoning quality

• Iterate on step guidance

3. Prompt Optimization

• Establish baseline metrics

• Identify failure patterns

• Adjust instructions for clarity

• Add/modify examples

• Tune output constraints

• 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