Prompt Engineering Guide

Topics covered

• •Understand what prompt engineering is and why it matters to researchers and developers.
• •Learn the key LLM parameters: temperature, top_p, maximum length, stop sequences, and frequency/presence penalties.
• •Break down the four elements of a prompt and the basic formats (zero-shot, few-shot, chat format).

Topics covered

• •Distinguish zero-shot from few-shot prompting and understand when each is enough.
• •Learn Chain-of-Thought (CoT) and Zero-shot CoT for reasoning tasks.
• •Understand Self-Consistency as a way to make CoT more reliable.

Topics covered

• •Break complex tasks into prompt chains (Prompt Chaining).
• •Understand the Tree of Thoughts (ToT) principle: searching a tree of intermediate 'thoughts.'
• •Know Reflexion — an agent improving itself through verbal feedback.

Topics covered

• •Understand RAG (Retrieval Augmented Generation) and its role in fighting hallucinations.
• •Learn the ReAct pattern: alternating reasoning and actions with external tools.
• •Know PAL and ART — delegating computation to an interpreter and to tools.

Topics covered

• •Generate synthetic data for training and testing using an LLM.
• •Use an LLM for code generation, completion, explanation, and debugging.
• •Understand the function-calling mechanism and its role in integrating LLMs with external systems.

Topics covered

• •Understand how instruction tuning (Flan) and RLHF (ChatGPT) shaped today's chat models.
• •Learn the specifics of prompting ChatGPT and GPT-4 (the system/user/assistant roles, steerability).
• •Get oriented in the open-model landscape: Llama, Code Llama, Mistral 7B, Mixtral, Phi-2, OLMo.

Topics covered

• •Recognize types of prompt attacks: injection, prompt leaking, jailbreaking.
• •Know defense tactics: cautionary instructions, parameterization, formatting, and a detector model.
• •Understand the problem of factual reliability (hallucinations) and ways to reduce it.