Prompt Engineering Guide - From Beginner to Advanced

Everything You Need to Know About Prompt Engineering

Introduction to Prompt Engineering

• The video introduces prompt engineering as a set of strategies designed to optimize interactions with artificial intelligence models.

• It explains that when using models like ChatGPT, Gemini, or Claude, users input natural language prompts and receive outputs in the same format.

• The effectiveness of these models relies heavily on how well prompts are structured, including word choice and examples provided.

Understanding Large Language Models (LLMs)

• LLMs function as prediction engines that generate text based on sequential input; they predict the next token (word or part of a word).

• A clear definition of prompt engineering is provided: it involves crafting high-quality prompts that guide LLMs toward producing accurate outputs.

• The speaker emphasizes the importance of adjusting settings for different LLMs to maximize prompt effectiveness.

Key Settings in Prompt Engineering

Output Length

• Output length determines the maximum number of tokens an LLM can produce in response; longer lengths allow for more detailed responses.

• Shortening output length does not necessarily lead to more concise answers; it simply limits how much text is generated before stopping.

Sampling Controls

• Sampling controls influence how LLMs select the next token by assigning probabilities to each possible token in their vocabulary.

Temperature Setting

• Temperature affects randomness in token selection: higher values yield more creative responses while lower values result in more predictable outputs.

Practical Examples and Demonstrations

Impact of Output Length on Responses

• An example demonstrates varying output lengths using a story prompt about a panda bear, showing how shorter limits lead to incomplete narratives.

Exploring Temperature Variations

Understanding Temperature, Top K, and Top P in AI Models

Temperature Settings

• The temperature setting influences the variability of responses from AI models; higher temperatures yield diverse outputs while lower temperatures produce consistent results.

• Adjusting the temperature is crucial depending on the use case; for creative tasks, a higher temperature is recommended.

Top K and Top P Sampling

• Top K sampling selects the top K most likely tokens based on predicted probabilities, affecting creativity in model outputs.

• A higher value for top K increases output variety, while a lower value leads to more factual responses.

• Top P sampling limits vocabulary selection based on cumulative probability; however, it is less frequently used by some practitioners.

Suggested Starting Points for Settings

• Recommended starting points include a temperature of 0.2 (lower than typical), top P at 0.95, and top K at 30 for coherent yet moderately creative results.

• To enhance creativity, increase both top P and top K settings; conversely, decrease them for more consistent outputs.

Prompting Techniques: Zero Shot vs. Few Shot

Zero Shot Prompting

• Zero shot prompting involves providing no examples to the model; only a clear task description is given.

• This method works well for simple tasks where extensive examples are unnecessary.

Example of Zero Shot Prompting

• An example prompt asks the model to classify movie reviews into positive, neutral, or negative sentiments without prior examples.

One Shot and Few Shot Prompting

• One shot prompting provides one example to guide the model's response; few shot uses two or more examples to establish patterns.

• More complex tasks benefit from additional examples as they help clarify desired output formats.

Example of Few Shot Prompting

• For parsing pizza orders into JSON format, providing structured examples ensures consistency in output structure across multiple requests.

Contextual and Role-Based Prompting

System Message Contextual Prompting

• System messages set overall context and purpose for language models by defining their roles (e.g., translating languages or classifying reviews).

Understanding Prompting Techniques in AI

Overview of System Instructions and Contextual Prompting

• The Google AI Studio features a clipboard labeled "system instructions," which is synonymous with the system message. This section allows users to set optional tone and style instructions for the model.

• Contextual prompting provides essential background information relevant to the task, helping the model understand nuances and tailor responses effectively. An example given involves writing for a blog about retro 80s arcade video games.

• The distinction between context and task is highlighted; while the task may not explicitly state the theme, contextual details guide the model's output.

Role Prompting: Assigning Identity to Models

• Role prompting assigns a specific character or identity to the language model, enhancing response consistency with that role's knowledge and behavior. This technique is notably effective in frameworks like Crew AI.

• An example illustrates role prompting where a user asks for travel suggestions as if they were a travel guide, demonstrating how this method can yield tailored recommendations based on location.

Step Back Prompting: Enhancing Insightful Responses

• Step back prompting encourages models to first consider broader questions related to specific tasks before generating answers, activating relevant background knowledge.

• This technique promotes critical thinking within LLMs (Large Language Models), allowing them to apply their knowledge creatively rather than responding generically.

Practical Application of Step Back Prompting

• A default prompt example shows standard prompting yielding generic results when asking for creative content like storylines in video games.

• By employing step back prompting, users can ask broader questions first—such as identifying key settings in first-person shooter games—before narrowing down to specific tasks like storyline creation.

Chain of Thought: Improving Model Outputs

• Chain of thought reasoning has been integrated into many models today, allowing them to articulate their thinking process before delivering final outputs.

Understanding Prompting Techniques in Language Models

Higher Quality Outputs from Models

• Many models come with built-in capabilities for generating higher quality outputs, but not all. Smaller or older models can still benefit significantly from effective prompting methods.

• An example using the Gemini 2.0 flashlight model illustrates how a non-thinking model can solve age-related problems correctly when prompted to think step by step.

Effective Problem Solving with Prompts

• Smaller or older models should be utilized for specific use cases where inference speed and cost are critical considerations.

• Combining different prompting techniques, such as one-shot or few-shot prompts with chain of thought, enhances problem-solving capabilities.

Chain of Thought Technique

• The chain of thought method allows models to mimic human-like reasoning by breaking down problems into smaller steps, improving accuracy in responses.

• This technique is particularly effective in STEM fields and logical reasoning tasks, leading to better outputs across various categories.

Self-Consistency as a Prompting Strategy

• Self-consistency addresses limitations in reasoning by combining sampling and majority voting to generate diverse paths and select the most consistent answer.

• By running the same prompt multiple times (e.g., five), the model can vote on which response it considers best, enhancing accuracy and coherence.

Practical Example: Email Classification

• In classifying emails as important or not, self-consistency was demonstrated through multiple outputs assessing an email about a bug report.

• Two out of three attempts classified the email as important based on its potential impact; however, this approach incurs high costs due to increased latency from repeated prompts.

Exploring Tree of Thoughts

• The tree of thoughts technique enables language models to explore multiple reasoning paths simultaneously rather than following a single linear chain.

Tree of Thought and React: Enhancing LLM Capabilities

Understanding Tree of Thought

• The process begins with multiple initial steps, where the model evaluates which is most accurate before proceeding to subsequent iterations until a final output is achieved.

• Implementing Tree of Thought effectively requires coding or a framework, as it becomes unmanageable when relying solely on user prompts for complex tasks.

Introduction to React Prompting

• React prompting combines natural language reasoning with external tools (e.g., search engines, code interpreters) to enable large language models (LLMs) to tackle complex tasks.

• This method mimics human operation by integrating logic with various tools that allow the model to gather knowledge, save memories, or communicate with other agents.

The Process of React Prompting

• The LLM first reasons about a problem and formulates an action plan before executing it and observing the results. This thought-action loop enhances task execution.

• Cutting-edge models like Gemini 2.5 Pro offer built-in tools for structured output and code execution but come at higher costs and latency compared to smaller models.

Practical Application of React Framework

• An example using Python code demonstrates how an LLM can be programmed to search for information about Metallica band members' children through a series of searches.

• Utilizing existing frameworks like Chain or Crew AI can simplify the development process without needing to create an agent framework from scratch.

Automatic Prompt Engineering

• Automatic prompt engineering allows users to generate detailed prompts without extensive manual input; this involves asking the model for a PRD based on brief descriptions.

• By feeding back generated PRDs into another model, users can streamline the coding process while ensuring comprehensive detail in their requests.

Deciding Between Code Execution and Natural Language Solutions

• A technique discussed involves determining when it's more effective for the model to write and execute code versus providing direct answers in natural language.

Best Practices for Prompting AI Models

Effective Code Execution and Output Accuracy

• The speaker discusses executing code and analyzing its output, emphasizing the importance of accurate results when prompting AI models.

• This process is referred to as "prompting using code," highlighting a method to ensure reliability in outputs.

Simplifying Prompts for Consistency

• It is recommended to provide examples (zero shot, one shot, few shot) when possible to achieve consistent outputs from the model.

• The speaker advocates for simplicity in design, suggesting that prompts should be straightforward and only include additional instructions when necessary.

Specificity in Expected Outputs

• Being specific about expected outputs is crucial; if a certain format like JSON is required, it should be clearly stated.

• The speaker emphasizes that vague requests lead to guesswork by the model, which can result in inaccurate responses.

Instructions vs. Constraints

• Clear instructions are preferred over constraints; instead of stating what not to do, focus on what actions or formats are desired.

Managing Token Length and Using Variables

• Controlling maximum token length is important for optimizing latency and cost in high-scale production scenarios.

• Utilizing variables within prompts allows for dynamic content generation; an example provided involves inserting city names into travel-related prompts.

Staying Informed About Model Capabilities

• Keeping up-to-date with AI model capabilities and limitations helps users format their prompts effectively.