LLMs and Prompting

Today LLMs can today perform extremely complex tasks like research, analysis, summarization, translation, code generation, advanced reasoning, creative writings and many more complex tasks - far exceeding todays search which is kind of information retrieval. LLMs are designed to be very intuitive with natural language interaction. Question to ask is are we getting the best optimal results out of these LLMs?

• Are people frustrated, confused, not satisfied with the output or results of LLMs? regardless if we use ChatGPT, Claude, Gemini, Grok.
• Are there specific LLM's for different types of real-world problems, scenarios and for different professions?
• for example is ChatGPT good for research and Analysis, where as, Claude could be brilliant in creativity?
• Are we clearly defining the task we want LLMs to perform?
• When do I use a chat Model LLM vs a Reasoning Model LLM?
• Are we clearly articulating the expectations of the details and/or end result we need?

There are many questions people have on the LLMs capabilities, and the answer simply lies in prompting, prompting techniques which is how we communicate with LLMs. We could call it providing commands, instructions, guidance or providing a detailed context depending on the output we want from LLMs.

There are many different types of prompting depending on what we want to get out of LLMs.

here are some highlights of my learning on prompting and prompting techniques;

1. System Level Prompting - the idea here is providing instructions to LLMs their role, capabilities, and the general rule of engagement, and we separate that from the real task prompt.
2. Markdown Prompting - as the name implies, we use the Markdown syntax (like # for titles, ## for subtitles, * or - for bullet points) within the prompt to explicitly request the LLM a structured output.
3. Style Prompting - we ask the LLM to explicitly provide the desired output with the writing style, tone, or voice. This is very useful for creative tasks, marketing. for example you could say for Style, we want it to be persuasive, or informative or creative and for Tone we could ask the LLM to provide the output as Professional, friendly, formal, Humorous etc.
4. Zero-Shot vs. Few-Shot Prompting - meaning for Zero-Shot we ask the LLM to perform a task without providing any examples of input-output pairs which essentially means we are relying solely on LLMs pre-trained knowledge. in Few-Shot we provide a small sequence of numbers, typically 1-5 of examples within the prompt to guide the LLM on the desired format or task execution. for example in Few-Shot prompt, learning a new concept is easier with few examples.
5. Chain-of-Thought (CoT) Prompting - we ask the LMM to breakdown its reasoning process into intermediate steps before arriving at the final answer and in such scenarios we could be adding in our input phrases like "Lets think step-by-step" or "Explain your reasoning" and the output performance improves on tasks which requires complex reasoning, arithmetic.
6. Skeleton-of-Thought (SoT) Prompting - this is a variation of CoT above where the LLM is first asked to generate an outline of "skeleton", for example bullet points of main topics of the response, and then elaborate on each point in the skeleton. a good analogy here could be if we are writing an article, get the article outline before the detailed full text.
7. Generated Knowledge Prompting - will instruct the LLM to first generate relevant background facts or knowledge related to the query before attempting to answer the main question which then serves as a context for the final response. output can be more in-depth answers.

There are many more types of prompting techniques which I will bring in the next round of my article.

In summary we need to get the best out of LLMs by practicing the prompt techniques which best provide the results on the context scenarios we are trying to achieve. To become good at prompting requires experimentation and developing a good "feel" for what works through real practice continuously.

Ultimately your PROMPT is the UI to getting the best from AI. Don't treat AI like search engine. ChatGPT/Claude/Gemini are the core LLM tools to master prompting.