prompt-engineer_skill

Overview

This skill helps you design, optimize, and evaluate prompts that get reliable, high-quality results from large language models. It focuses on practical techniques—few-shot, chain-of-thought, system prompts, and structured outputs—while balancing token cost, latency, and model-specific behavior. Use it to create reproducible prompt pipelines and measurable evaluation frameworks for production or experimentation.

How this skill works

I start by clarifying the task, success criteria, constraints, and edge cases. Then I choose an appropriate prompting pattern (zero-shot, few-shot, CoT, ReAct), author clear role-and-task instructions, and define structured output schemas when needed. I test prompts with diverse cases, measure accuracy and consistency, iterate to reduce tokens or fix failure modes, and deliver versioned prompts with test results and deployment guidance.

When to use it

• Designing prompts for a new LLM-powered feature or agent
• Optimizing an existing prompt for accuracy, cost, or latency
• Implementing advanced techniques like chain-of-thought or few-shot learning
• Creating system prompts with personas, guardrails, and context rules
• Building structured output formats (JSON mode, function calling) and validating schemas

Best practices

• Define clear success metrics and collect representative test cases including edge cases
• Use few-shot examples that match the target distribution and avoid contradictory examples
• Validate structured outputs against schemas and enforce parsing checks
• Measure cost and latency trade-offs; reduce tokens where possible without harming quality
• Version prompts, document behavior and limitations, and monitor for drift in production

Example use cases

• Create a diagnostic prompt for customer support triage that outputs a fixed JSON schema for downstream routing
• Tune a summarization prompt to reduce hallucinations and enforce length constraints across model versions
• Build an evaluation suite that compares prompt variants on accuracy, consistency, and token cost
• Migrate prompts between providers by adapting instructions to model-specific strengths and testing edge cases
• Implement chain-of-thought for complex multi-step reasoning and measure improvement on targeted benchmarks