prompt-engineering

This skill provides workflows to write, debug, and optimize prompts for large language models. It solves the recurring problem of ad-hoc, unreliable prompting by offering structured, testable, and reusable patterns for few-shot example selection, chain-of-thought reasoning, prompt optimization, template composition, and system prompt design.

The skill keeps a concise main workflow and defers deep guidance to targeted reference files loaded per pattern. It covers five core areas: few-shot learning (example count of roughly 3-5, semantic-similarity and diversity selection, progressive complexity ordering, with concrete templates for classification, transformation, extraction, and error correction); chain-of-thought reasoning (zero-shot and few-shot CoT, self-consistency by sampling multiple paths, and verification steps); prompt optimization (baseline measurement, single-variable changes, tracking accuracy, consistency, token efficiency, robustness, and safety, and A/B validation with statistical significance); template systems (clear variable naming, conditional sections, role-based and hierarchical composition); and system prompt design (role and expertise specification). Reference material includes reusable templates and example selection utilities such as a context-aware selection routine.

Target users are prompt engineers, application developers, and anyone building LLM-powered features who needs to draft, restructure, or measurably improve prompts and system prompts, add few-shot examples or reasoning scaffolds, or create reusable prompt templates. It is a provider-agnostic guidance skill focused on prompt design methodology and evaluation rather than any specific model API.