This repository aggregates instruction-based prompts designed for GPT-instruct (2020) and GPT-3.5-instruct, the first instruction-tuned GPT-3 variants. It covers prompt phrasing and positional sensitivity strategies—documenting how small wording changes can swing accuracy by up to 76 percentage points and explaining primacy/recency biases. The material focuses on optimizing zero-shot prompting for base and SFT-only language models.