Before ChatGPT, prompt engineering was just verbs and hope

What it does This repository is a curated phrasebook for GPT-instruct and GPT-3.5-instruct models. It catalogs specific instruction verbs—like Explain, Generate, and Classify—and provides copy-paste prompt templates for tasks such as text generation, classification, translation, and summarization. The author organizes these around a simple three-part structure: direct background, context, and action instruction.

The interesting bit Unlike many prompt-gardening guides, this one cites actual research on model behavior, noting that base and SFT-only models can swing by up to 76 percentage points depending on prompt formatting and suffer from a “Lost in the Middle” positional bias. It also candidly admits that modern RLHF-aligned models are “noticeably sturdier,” making the repo a useful time capsule of pre-chatbot prompt engineering.

Key highlights

• Curated lists of high-leverage instruction verbs for classification, generation, transformation, and comparison tasks.
• Ready-made prompt templates for programming, translation, data extraction, and text categorization.
• Explains the disproportionate weight carried by the first few tokens, especially verbs like Given and Following.
• Explicitly distinguishes between base/SFT models and RLHF models, clarifying where these sensitivities actually matter.

Caveats

• The README itself notes that RLHF-aligned models are far less sensitive to these tricks, so much of the advice is moot for modern ChatGPT users.
• The text contains visible typos and formatting inconsistencies, suggesting the project is not actively maintained.
• The prompt examples are rooted in the early GPT-3.5-instruct era and may feel obvious to anyone accustomed to chat-tuned models.

Verdict A handy reference if you are still hitting legacy instruct APIs or studying how early alignment worked. Everyone else can probably rely on their intuition with modern chat models.