Prompt Engineering

Giving the AI brain a unique set of instructions to increase its intelligence and responsiveness is what AI prompt engineering entails. To comprehend what we want from AI models like ChatGPT or GPT-4, they need to be gently nudged in the right direction. Prompt engineering can help with it. The finest answers from the AI may be ensured by carefully structuring the prompts. Now, prompt engineering doesn’t only happen once. The process of adjusting and experimenting is continuing. When we ask the AI a question, we experiment with varied wording and the addition of unique rules. We seem to be concocting a miraculous concoction of instructions! Let’s take a look at some rules to construct good prompts to generate accurate results for AI.

Some Rules of Prompt Engineering

Rule 1: Use latest AI models for your tasks

It is important to use latest AI models, because AI models do not tend to give accurate results when they are not updated with recent data. For example, ChatGPT was trained on data dated till September 2021 so it won’t be able to provide you with information after September 2021. Whereas, GPT-4 (or ChatGPT plus) is trained on latest data and is able to generate the latest information.

Rule 2: Start your prompt with Instructions, and use “”” or ### to separate your input from your instruction.

For example,

Translate the below text to Hindi.
Text: """
I work at Geeks for Geeks
"""

Rule 3: Be as precise, illustrative, and thorough as you can when describing the context, result, length, format, style, and other desired characteristics.

For example,

Write a short inspiring and funny poem about Geeks for Geeks, focusing on their recent launch of Full-stack development course in the style of Sandeep Jain.

Rule 4: Try including examples to illustrate the desired output format.

For example,

Create a list of datasets that can be used to train a logistic regression model. 
Ensure that the datasets are available in CSV format. The objective is to use this dataset to learn about Supervised Learning. 
Also, provide links to the dataset if possible.
Desired Output format:
Dataset name | URL | dataset description
             |     |

Rule 5: Start with zero-shot learning, then few-shot learning. If neither of them worked, then fine-tune.

Zero-shot: The ability to complete a task without being trained for it.

Few-shot: The ability to complete a task after being trained with a limited number of examples.

Fine-tune: By fine-tuning GPT-3 using training data, you can increase its performance on particular tasks. When the model is being fine-tuned, it reads text tokens from the training set and predicts which token will come next. The model’s internal weights are updated when it predicts incorrectly to increase the likelihood that it will forecast correctly the following time. The model will eventually learn to generate the pattern of tokens shown in your training data after receiving sufficient training.

For example,

Zero-Shot:

Find the odd one out from the following:
Text: """Fish, Octopus, Crab, Lion"""
Answer: 

Few-Shot:

Find the odd one out from the following:
Text: """Fish, Octopus, Crab, Lion"""
Answer: Lion
Text: """Ostrich, Cheetah, Turtle"""
Answer: Turtle
Text: """square, rectangle, rhombus, circle"""
Answer:

Fine-Tune: You can see the best practices for fine-tuning here.

Rule 6: Don’t add imprecise descriptions to your prompt

For example,

Write a paragraph of 3 to 5 lines explaining apple.

Rule 7: Instead of mentioning “what not to do”, try including “what to do” to your prompt

For example,

Write a conversation between a father and a son, the conversation should be about future goals and refrain from involving any personal information. 
Instead of involving personal information, try directing the conversation towards college.

Rule 8: Specifically for code generation: Use “leading words” to guide the model towards a certain pattern.

For example,

# Write a simple python function that
# 1. Ask me for an integer number 
# 2. Generate a fibonacci series till the input number using top-down approach
import

The “import” instruction tells the model in the code snippet below to begin writing in Python. (Likewise, “SELECT” is a good starting point for a SQL statement.)

You can use these rules to design some prompts and execute it on ChatGPT. You can see how different the results are from your regular prompts.

OpenAI is the leading company in the field of AI. With the public release of software like ChatGPT, DALL-E, GPT-3, and Whisper, the company has taken the entire AI industry by storm. Everyone has incorporated ChatGPT to do their work more efficiently and those who failed to do so have lost their jobs. The age of AI has started and people not adapting to AI could introduce some difficulties for them. 

In this article, we will be discussing how you can leverage the power of AI and make your day-to-day tasks a lot easier by using the OpenAI APIs (Application Programming Interface) that allow developers to easily access their AI models and Integrate them into their own applications using Python.