Prompting secrets

Chat GPT and all other AI chat models are here to stay and be part of our work culture so instead of fighting it bell the cat and make it dance to your tuneFor the uninitiated prompt engineering refers to the art of crafting effective prompts to get the desired responses from the chat model This is a crucial skill because Large Language models are trained on a variety of data sets both real and now synthetic and need clear logical prompts to give you the desired informationThis is a skill that can be learned it requires little or no prior knowledge and your skill will improve the more you practise Using prompt engineering will also help you extract the most out of ChatGPT regardless of if you plan to use it for school or office work But before we know how to game the system we have to know a little bit about how it all worksWhat is LLM and how it worksA Large Language Model LLM is a type of AI designed to understand and generate humanlike text based on the patterns it has learned from massive amounts of language data It works by breaking down text into tokens small units of language like words or subwords and analysing the relationships between them When you give an input the model predicts the most likely next words or sentences by using patterns and statistical relationships it learned during training Essentially it ldquocompletesrdquo your input while mimicking natural language allowing it to respond to questions write text translate languages etc This is also the reason that you should not blindly trust anything that comes out of an LLMIt is merely in simplistic terms predicting the next likely word like autocorrect does So how do we get the right information from the LLM With a good prompt of courseWhat makes a good promptFour Cs that make a good prompt Content Clarity Context and ConstraintsContent Go straight to the information you need Use direct language and include key details If you are looking for a particular format for the answer include that For example ldquoCreate a table with the outputrdquo or you could provide some samples and ask the LLM to mimic the style of the response to that style You can invoke personas to improve your output but more on that laterClarity It is very important to know what you are looking for and be able to explain clearly to the LLM what you want Even a little domain knowledge is important for this step For example if you are using ChatGPT to write some Python a programming language code for yourself you will need to know at least the basic functions or have a basic understanding of how Python works Having this knowledge will give you the ability to write a much more detailed prompt and get results faster Context LLMs have been trained on a huge amount of data that they use to predict text Some of this data may be irrelevant So always give the LLM contextFor example write an essay about AI would not have any context it will generate things that might not be in the scope of your essay Instead if you give it context by saying lsquoWrite an essay about use of AI in food marketingrsquo in 500 words using American slang Embellish your prompts as much as possible with all the details you can Adding documents and tables can enhance the output closer to desired outputConstraint There are several ways to constrain the output of ChatGPT An easy one is word count as in the case of the essay example A word count will make sure that the LLM generates the right amount of text needed for the assignment You can also give the LLM emotional constraints like mentioning that the essay is a significant part of your grade and that you need it to be the best essay about AI in the world Also please refer to the clarity point here because you need to have some domain knowledge to know that the LLM is not hallucinating the facts in the essayMake helpful PersonasPrompts also benefit from creating personas For example you need the same essay on AI Tell ChatGPT to pretend that it is the foremost expert on AI and that everything it says is backed by scientific papers This persona has a twofold benefit the first is that it will pretend to be that expert and the second is that it will now reference all it says against scientific papers and articles You can also ask it to cite sources for additional flair However this doesnrsquot protect the LLM from hallucinating and you will need some domain knowledge to crosscheck You can also use this persona to check your workPractise and play aroundPrompt Engineering is also about poking prodding and pushing the limits of the LLM So if you donrsquot get your desired result on the first try keep trying different approaches Eventually you will figure out the best way to work for your use case