Prompt Engineer

Problem It Solves

Prompts often fail because requirements are not decomposed into model-executable instructions: a system prompt mixes role, constraints, and examples; few-shot cases cover only ideal inputs; JSON output drifts occasionally; and behavior changes when switching models. This skill turns vague requirements into testable prompt designs, focusing on reducing output uncertainty rather than writing longer prompts.

How It Works

• System prompt design: splits role definition, constraints, output format, and examples, and chooses strategies for reasoning models versus general-purpose models.
• Prompt optimization: applies chain-of-thought, few-shot, reverse constraints, and format checks where needed, reducing reliance on implicit interpretation.
• Testing and iteration: builds cases for normal, empty, ambiguous, and boundary inputs, then tracks accuracy, consistency, and format compliance before deciding whether to add constraints, examples, or restructure the prompt.
• Deployment and monitoring: keeps version notes and regression tests, so behavior does not drift after model updates.

Scope and Guardrails

• Focuses on LLM system / user prompts, not image or video generation prompts.
• Does not publish automatically or rewrite local files silently; file writes request confirmation.
• Does not design prompts that bypass model safety controls, and sensitive scenarios require disclaimers.