In view of the huge capabilities of ChatGPT, it is imperative to learn ChatGPT usage skills in depth. As a large language model (LLM) such as ChatGPT, there is also a new engineering field: prompt engineering (Prompt Engineering).
Prompt Engineering (Prompt Engineering) is a relatively new discipline, which focuses on the development and optimization of prompt words, and helps users use Large Language Model (Large Language Model, LLM) in various scenarios and research fields. Having skills related to hint engineering will help users better understand the capabilities and limitations of large language models.
Cue engineering is not just about designing and developing cue words. It encompasses various skills and techniques for interacting with and developing large language models. Hint engineering plays an important role in realizing the interaction and docking with large language models, as well as the ability to understand large language models. Users can improve the security of large language models through hint engineering, and can also empower large language models, such as using professional domain knowledge and external tools to enhance the capabilities of large language models.
Translated into vernacular: Technology developers have more abilities to mine AI (code writing, API calls, etc.), if ordinary people also master these skills, they can dig out greater ChatGPT capabilities than ordinary users.
prompt word
text summary
information extraction
question and answer
Text Categorization
dialogue
code generation
reasoning
Here are a few special prompts to choose, and focus on talking about them.
Information extraction , using this command allows you to find suitable targets from large blocks of text
Reference instruction: instruction+###content###
Text classification , classify according to different categories, or make classification output based on certain sample references.
Reference instruction 1: For which aspects of the content, classify +%%%content%%%
Reference instruction 2: instruction
Reasoning , currently the reasoning task is the most challenging for large language models. The most exciting thing about inference tasks is that it can lead to various complex applications born from large language models. The big language model is good at word processing, and the logic of mathematical reasoning is indeed a bit difficult, but it can be handled in general.
Here's a slightly more complicated command:
command prompt
Zero Sample Prompt
Few Sample Hints
Chain of Things (CoT) Thinking Prompts
self-consistency prompt
generate data tips
With the instructions, sometimes it is necessary to add some style hints, so that ChatGPT can work more accurately.
1. Zero style hints, work without any hints
2. Small sample tips, provide some cases for ChatGPT to learn, and then output content based on the examples
3. Chain-of-Thought Prompting, the full text is called: Chain-of-Thought Prompting , suggesting that ChatGPT realizes complex reasoning ability through intermediate reasoning steps. You can combine this with few-shot hinting to get better results for more complex tasks of reasoning before answering. 
Chain thinking is the key skill. For some slightly complicated questions, ChatGPT’s performance is not satisfactory. You need to instruct ChatGPT to follow certain steps to solve the questions step by step, and the final answer is often correct. 
Is it a bit of nonsense, let's adjust the instructions, and the answer at this time is correct. 
4. Self-consistency hints, the idea is to sample multiple different inference paths via few-shot CoT and use the generated results to select the most consistent answer. This helps improve the performance of CoT hints in tasks involving arithmetic and commonsense reasoning. 
(It’s also nonsense, sometimes mistakes will be made in one step, let’s let ChatGPT take it slowly)
5. Generate data prompts, generate data in a specified format according to certain data prompts, instead of being freely played by ChatGPT. 
To sum up, it is a relatively common technique for advanced use of ChatGPT. There are also some more advanced techniques that may involve some coding skills.
Further reading: