Tips for prompting based on LLM

Feature / Style	OpenAI (GPT-4, GPT-3.5)	LLaMA (LLaMA 2 & 3)	Mistral (Mistral 7B, Mixtral)
Role Structure	Uses system, user, and assistant roles explicitly	No native role handling, but you can simulate it using text cues	No role handling, but follows text cues and prompt templates
System Prompt Support	✅ Full support — allows defining model behavior (e.g., "You are a helpful assistant.")	🚫 No native support; must embed in user prompt manually	🚫 No native support; simulate via prompt prefix
Formatting Style	Natural conversation, JSON-compatible, markdown-friendly	Structured, requires consistent formatting for few-shot and instruct modes	Concise and direct; works well with bullet lists, steps, or templates
Few-shot Learning	Highly effective with few-shot examples	Effective, especially with CodeLLaMA and LLaMA-Instruct variants	Can benefit from few-shot, though prefers minimal examples
Chain-of-Thought Reasoning	Strong performance with "Let's think step by step" style prompts	Improves performance significantly with explicit CoT instructions	Supports CoT well, especially in instruct-tuned variants
Prompt Length Handling	Handles long prompts well (especially GPT-4-1 with large context windows)	Medium capacity; recent models like LLaMA 3 support longer prompts	Smaller context (e.g., 32K tokens), favors concise prompts
Fine-tuning Response Format	Easily aligns to JSON, tables, and multi-part instructions	Needs more specificity to get consistent formatting	Consistent if given strict format constraints
Use of Delimiters	Often uses """ or ### to separate instructions from input	Suggested to separate examples and instructions clearly	Benefits from template-like structures, including consistent line breaks
Multimodal Input Handling	GPT-4o supports images and audio	LLaMA 3 (future) may add modalities; current LLaMA is text-only	Mistral is text-only for now