Prompts•Diag_Ele•4dOpinion

Some Methodological Thoughts on Prompting

As shown in the diagram: E: Textualization or encoding mapping; C: Channel transmission relationship; ideally, C(X)=X; D: The recipient’s decoding and contextual inference process; Ne: General disturbances such as omissions and ambiguities during the expression stage; Nc: Technical noise such as typos, garbled text, and truncation; Nd: Comprehension biases caused by proper-noun ambiguities, missing context, and so on. In general, a person’s original information passes through a series of processes and is mixed with noise, so the recipient’s final understanding by AI will contain deviations. The purpose of prompting is to reduce noise in this process by minimizing omissions and ambiguities, so that the original information intended by the human brain is as close as possible to the recipient’s final understanding by AI. Of course, this model can also work in reverse. AI’s original information is conveyed to humans through language and text, but because of limits on the number of words generated, a human’s theoretical knowledge base (manifesting as missing context) can also lead to comprehension biases. Therefore, based on the above theory, I have come to the following reflections: 1. A prompt is not necessarily the better the longer it is. If you only describe a single turn, this amounts to open-loop control, and you may have no way of knowing how much of a discrepancy exists between the recipient’s understanding by AI and your own understanding. A prompt of an appropriate length can ask the recipient to restate and pose questions, allowing you to discover the recipient’s final understanding by AI and then fine-tune it accordingly, creating a feedback closed-loop control that improves control precision. This is what is known as context engineering. 2. To improve the precision of AI control, you need not only an AI model capable of correctly expressing its understanding, but also a human capable of understanding the text produced by AI. Humans therefore need a basic foundation of knowledge in the relevant industry; otherwise, enormous noise will be introduced into the communication process with AI, making precise control impossible. 3. If cross-domain operations are required—for example, writing code for personnel statistics as someone working in human resources—you first need to get on the same wavelength with AI and ensure that the human can correctly understand AI’s expression. There are two ways to do this: first, let AI know what kind of person it is communicating with (a human resources professional), so that AI voluntarily avoids using technical terms from the computer field, trading some loss in expression precision for greater precision in human understanding; second, actively align with AI’s expression by proactively learning knowledge from the computer field. These are my personal views and insights. Please feel free to point out any improper expressions or errors, and I welcome corrections, additions, and discussion.