Understanding Prompt Engineering

Prompting Unveiled

• The output may not be perfect on the first try. It may have errors. Instead of criticizing the output, think about the additional context you can provide to the model to improve the output on the next iteration.
• A basic prompt leads to a basic answer. A well thought out prompt with rich context leads to a more accurate answer.
• Every day we use prompts without even realizing it. When you search on Google, question Siri, or ask a friend for advice, you’re using prompts.
• A prompt isn’t just a question. It’s carefully crafted, with each word and phrase chosen to elicit the desired response.
• An effective prompt is :
  • Clear - Provide relevant context to reduce ambiguity
  • Specific - Avoid unnecessary information to reduce noise and get close to the answer
  • Open-ended - Surface non-obvious insights to allow the model to think outside the box
• Here an example rassembling the three elements:

I'm an 8th-grade math teacher preparing to teach trigonometry.
Design a lesson plan tailored for a 60-minute session with 20 students.
Can you also recommend some create, real-world examples to make the lesson more engaging for the students ?

• Prompt engineering is the art of writing inputs that maximimez the quality of the output.
• Masterting prompt engineering also involves understanding what not to do :
  • Overloading - Providing too much information can dilute the essence of your query
  • Ambiguity - Being vague can lead to generalized answers
  • Over-complication - Using jargon, complex language, or overly technical terms can confuse the model and lead to misinterpretations or overly complex outputs
• Prompts are not static. The beauty lies in their iterative nature. Feedback driven refinement enhances prompt quality, creating a continious loop of improvement. Think of it as a dialogue, with both user and model striving for perfection.
• When writing prompts, there are other stuff you can do to improve the result you are looking for :
  • Infusing persona
    • Adopting a persona is sometimes helpful to influence for example output style, audience, length and tone
  • Making it personal
    • Providing examples is a great way for the model to emulate your writing style
  • Training techniques
    • Zero-shot prompting - No structure provided
    • One / Few-shot prompting - Roadmap of steps to follow
• ChatGPT has its limitations. Understanding them is key to crafting effective prompts. Here are some of them :
  • Biases - The model presents stereotypes or misinformation
  • Hallucinations - The model confidently states incorrect information
  • Overfitting - The model is only as good as the data it’s tained on

Prompting Strategies and Techniques

• Crafting a prompt is like laying foundation for a building. It sets the stage for everything that follows.
• Just as a recap, remember a good prompt is composed of clarity, specificity, and openness.
• Here are the five elements that can turn a good prompt into a great one :
  • Instructions
    • Make your instructions clear to reduce ambiguity
  • Persona
    • Adopt a persona to influence output style, audience, length and tone
  • Output format
    • Specify the output format to ensure consistency and accuracy
  • Context
    • Provide relevant context to reduce ambiguity
  • Examples
    • Provide examples to guide the model’s output
• Harnessing the true potention of LLM like ChatGPT isn’t just about asking the right questions but also about steering the response in a desired direction. Think of yourself as the conductor of a vvast informational orchestra, guiding it to play the exact tune you want to hear.
• While ChatGPT knowledge base is impressive, to truly make the most of it, we need to understand how to guide its outputs. This is achievied using the SALT framework :
  • Style - Define the framework
  • Audience - Tailoring the content to the audience
  • Length - Controlling the length of the output
  • Tone - Influencing the output style and tone
• ChatGPT has a knowledge cutoff date. This refers to the point in time when the model was last updated with new data. Any new information after this date is not available to the model. Meaning that the model can’t answer questions about the latest events or information.
• Engaging ChatGPT yields a myriad of response but it’s vital to judge the quality of the output. For this, you can use the LARF framework to evaluate responses :
  • Logical consistency
  • Accuracy
  • Relevance
  • Factual correctness
• The way you present your prompt is just as important as the content within it. Here are some best practices to ensure clarity, precision and optimal resopnses :
  • Use markdown
    • Heading # to break up your prompt
    • Bold ** to highlight key words or phrases
    • Quotation " marks to indicate the start and end of a quote
    • Delimiters --- to separate different parts of your prompt

Advanced Prompt Engineering

• At the core of your interactions lie training techniques. These determine the way the model generates answers. It’s crucial to understand the spectrum of zero-shot, one-shot and few-shot learning. They represent the degree of examples or context we provide ChatGPT before asking our main question.
• One fascinating aspect of few-shot learning it that ChatGPT becomes more than just an autocomplete tool. It turns into a pattern matching and pattern generation engine.
• Chain of Thought (COT) is an advanced training technique that involves providing the model with a step-by-step reasoning process. This technique helps the model to think through the problem and generate a more accurate and detailed response.
• Every model, regardless of its sophistication, has its set of limitations. These limitations stems from the data it was trained on. By recognizing these limitations, we can better craft our prompts to avoid them.
• ChatGPT’s knowledge is almost one-dimensional, meaning that it’s not able to process complex concepts or abstract ideas. You have to ask question from a certain direction to get the answer you are looking for. We still don’t know exactly how language models work. What we do know is that LLM’s learn from vas amounts of data sourced from the internet. Consequently, they can inherit and sometimes amplify biases present in the data.
• Hallucinations in the context of LLM’s refer to instances when the model confidently states incorrect information. This is often due to the model’s inability to distinguish between real and fake information. Asking ChatGPT to provide sources of information often leads to correcting itself.
• Overfitting occurs when a model is too closely tailored to its training data, making it less effective at generalizing to new, unseen data. This can lead to inaccurate or irrelevant responses. The model currently sucks at reasoning and it’s not able to handle complex concepts or abstract ideas suck as writing jokes, developing scientific experiments or creating poetry.
• One brilliant feature of ChatGPTY is that within an individual conversation, each prompt build upon the previous response. This creates a rich dialogue that uncovers deeper insights and fosters a comprehensive understanding of the topic.
• Iteration is key to mastering prompt engineering. It’s not about getting the perfect prompt on the first try. It’s about refining your prompt through trial and error.
• Encouraging the model to engage in deliberate thinking helps not only prevent premature conclusions but also ensures the model fully comprehends the task at hand.
• When approaching complex tasks or eeking detailed outputs, breaking down the prompt into step-by-step instructions can help the model to generate a more accurate and detailed response.
• Push ChatGPT to think before presenting and answer. By instructing the model to deliberate or reason through a problem, you’r essentially giving it “thinking time”. This often results in a more accurate output as the model takes a moment to sift through the information and generate a more thoughtful response.
• Each response of ChatGPT is a feedback loop. It’s a chance to refine your prompt and improve the output. We can sharpen our prompts by asking ChatGPT to explain its reasoning, provide sources of information, or even ask it to correct its own mistakes.

All content written above is coming from the “Understanding Prompt Engineering” course on DataCamp.