prompt-engineer_skill

Overview

This skill is an expert prompt engineer for designing effective prompts for LLM-powered applications. It focuses on translating intent into reliable instructions, iterating prompts with measurement, and specifying output formats to reduce ambiguity. The skill emphasizes structure, context management, and systematic evaluation to maximize model reliability.

How this skill works

I inspect task goals, user intent, and expected outputs to create clear system prompts, few-shot examples, and constraints. I design prompt scaffolds (role, context, instructions, constraints, output format, examples), tune for token and context limits, and produce test cases to measure changes. I also recommend temperature and decoding settings and defend prompts against common injection and ambiguity failures.

When to use it

• Creating or refining system prompts for Claude or other LLMs
• Designing few-shot examples to teach desired behavior
• Specifying exact output formats for downstream parsing
• Debugging inconsistent or hallucinating model responses
• Optimizing prompts for constrained context windows

Best practices

• Structure prompts into Role, Context, Instructions, Constraints, Output Format, and Examples
• Use 2–5 diverse few-shot examples including edge cases and negative examples
• Measure the impact of each change with A/B tests or unit-style prompt tests
• Limit context to relevant facts and curate examples to avoid bias
• Specify exact formatting (JSON, CSV, bullets) and validate parsed outputs

Example use cases

• Build a system prompt that enforces a legal-style checklist and returns JSON
• Create few-shot examples to teach a model domain-specific summarization
• Design chain-of-thought prompts for multi-step reasoning with separate reasoning output
• Tune prompts for a CLI that configures and monitors Claude Code with strict output parsing
• Develop defensive prompts that sanitize user inputs to mitigate prompt injection