Currently, there is no truly user-oriented AI App, while ordinary people hope to use AI to help them complete some basic tasks. Actually, there are some methods that can be referenced.
Next, I will synthesize the content I have written before, using writing a piece of copy as an example, to organize a step-by-step operation guide, hoping to bring some help to those who read this article.
This is a method for using AI to create copy, a method for product manager requirement design, and also a reference approach for designing Agents.
Points to note:
- This is a methodology designed around the current defects of AI models and conversational interaction.
- It is not the user's fault if they cannot ask questions to AI; it is a product defect.
- When using AI, always be vigilant about the content it outputs.
- Prevent the accumulation of errors in multi-turn dialogues.
- Use AI well, rather than being limited by AI.
I believe that as long as you use it more, everyone can summarize a methodology that suits them. Mastery comes with practice for any tool; the more you use it, the more naturally you will integrate and understand it.
Step 1 Define Roles: Who "You" Are, Who "It" Is
"Who it is": This tells AI what identity to adopt to solve the problem. Narrow the abstract large model into a professional role to reduce irrelevant answers.
A chef and a firefighter have different attitudes toward fire. When solving the problem of "the stove catching fire," the chef will lift the pot, while the firefighter will pull out the fire hose.
Making AI clearly understand its role can effectively constrain its reasoning patterns, facilitate its use of built-in expert models, and prevent it from using a fire hose to cook.
"Who I am": That is, what identity the user has, telling AI who it is working for.
Again using the chef and firefighter as examples, whether "I" am a diner or a citizen affected by the fire will also influence AI's output.
If "it" is a chef and "I" am a citizen affected by the fire, it might use a wok to help me put out the fire. Although current AI already has certain judgment capabilities, it is not guaranteed that such problems will not occur in some specific scenarios.
"You are a writer proficient in English writing, and I am a high school senior student."
Step 2 Define the Scenario
Defining the scenario is telling AI the environmental information you are in and the information you have prepared to solve the problem. This prevents AI from fabricating conditions and makes the output fit the real situation.
For a chef, working on a home stove is different from working in a commercial kitchen. Moreover, you need to tell it what ingredients you have that can be used for cooking. After all, if basic conditions are lacking, this AI chef might go off on its own. For AI, lacking conditions means it will fabricate them; as long as it is logically consistent, it does not care about authenticity or usability.
You need to clearly tell AI about your working environment, your state, your materials, etc. This information is the foundational condition for solving the requirement. For example:
"I am writing an English article about an outing, and during the outing, I did xxxxxxx."
Step 3 Define the Requirement
Defining the requirement means clearly informing AI what work you need it to complete for you, which is also AI's work objective. Clarifying the goal avoids off-topic output.
This AI chef can effortlessly handle all eight major Chinese cuisines, but which cuisine it will cook depends on which cuisines appear frequently in its training data. Therefore, you need to define the work objective for AI, so that if you want it to cook Shandong cuisine, it only cooks Shandong cuisine; if you want Cantonese cuisine, it only cooks Cantonese cuisine.
"I need you to write this English article for me" rather than teaching me how to write an article.
Step 4 Define the Framework
This involves defining the structure of the content AI outputs and also putting forward specific requirements for AI. It makes the result closer to a usable form rather than a pile of scattered text.
It is like "ordering dishes." Asking the chef to cook Cantonese cuisine is the requirement; "the first dish is steamed fish, the second dish is xxx" is ordering. Making AI output according to the form you envision is a key step for structured output.
At the same time, you can also make stylistic requirements to AI, such as salty, sweet, bitter, or spicy. For example:
"I hope the first paragraph expresses my feelings about going on an outing with friends, the second paragraph writes about interesting things on the bus, the third paragraph xxxxxxx. I want the article style to be lively, with a word count within 1000 words xxxxxxx."
If you are unsure how to define the structure, you might as well send the first 3 steps to AI and let it design based on your requirements.
"Please help me design the structure of this English article and provide style recommendations."
Step 5 Define the Content of Each Module
In writing, this means controlling each paragraph, which can be understood as more precise control over AI's output content. This step can prevent AI from losing control in long tasks and also facilitate local corrections.
If you ask a chef to make steamed fish for you, the chef will only make it according to its own taste. If you have requirements for how the fish is cooked, you can clearly ask the chef to make a dish that suits your taste; you can ask for vinegar in this steamed fish, or specify whether to use grass carp or sea bass, etc.
More precise and clear definitions can help you better control AI to meet your expectations. If you need to write an article more complex than a high school English essay or complete a more complex task, this step is essential.
For example, when writing an article, you can propose "In the first paragraph, I hope to express my joyful and excited mood because I will meet my long-unseen friend," or if you have an article link for reference, you can also do this: "The first paragraph can refer to this article: https://www.xxxx.com."
If the content you need to construct is relatively complex or you have a lot of materials to provide, it is best to start a new conversation for each module's definition. Send the first 4 steps and the module you need to create to AI and define them separately. For example:
"You are a writer proficient in English writing, and I am a high school senior student.
I am writing an English article about an outing, and during the outing, I did xxxxxxx.
I need you to write this English article for me.
I hope the first paragraph expresses my feelings about going on an outing with friends, the second paragraph writes about interesting things on the bus, the third paragraph xxxxxxx.
In the first paragraph, I hope to express my joyful and excited mood because I will meet my long-unseen friend.
Please help me write the content of the first paragraph."
After completion, start a new conversation and repeat this process, continuing with "Please help me write the content of the second paragraph." Some AI might presumptuously complete all modules at once, which will inevitably cause information confusion in complex long texts. To avoid this situation, you can add one more sentence at the end:
"Only complete the task according to my requirements; do not do anything extra."
Step 6 Module Combination
Effectively combining each module is equivalent to plating and serving dishes. In a banquet, each dish seems to stand alone, but a chef with aspirations will control the connection between dishes. Specifically in writing, it is the transition and connection between paragraphs.
If the article is output by AI at once, this step can be ignored. If it is a more complex creation, you need to merge the information from steps 1, 2, 3, and 5 and send it to AI, asking it to complete the integration.
"You are a writer proficient in English writing, and I am a high school senior student.
I am writing an English article about an outing, and during the outing, I did xxxxxxx.
[First paragraph text]
[Second paragraph text]
……
I need you to complete the integration for me, requiring logical coherence and tight connections." At this point, your requirement for AI is no longer to write an article but to combine existing content, which is optimization.
Step 7 Proofreading and Checking
This is the final step; you need to check the content output by AI. It is equivalent to tasting the dishes made by the chef. This step is crucial because it prevents hallucinations and errors, ensuring reliable results.
When using AI, you need to always be vigilant about the content it outputs. Because no matter how you constrain and require AI, hallucinations are an inevitable problem at this stage. It will fabricate facts, presumptuously modify your requirements, and misinterpret the meaning of materials. It will mistake sugar for salt and also deceive you by calling shrimp fish.
If you really cannot determine whether the content output by AI is correct, you can ask it to provide citation sources or require it to perform semantic comparisons on certain content and references.
"Check if the article meets my requirements."
"Conduct fact-checking and list information sources."
"Compare information sources and check for erroneous descriptions."
At this point, you can have AI complete the creation of an article for you.