iTranslated by AI
My ChatGPT Custom Instructions Setup
Custom Instructions are information that can be passed to all chats in common.
It is convenient because you no longer need to pass prerequisite information or desired responses for each chat.
At the time of publication, there is no way to switch between multiple Custom Instructions, so it seems best to provide generic instructions that can be used for any situation.
What would you like ChatGPT to know about you to provide better responses?
The tips on the Custom Instructions setting screen are as follows:
What would you like ChatGPT to know about you to provide better responses?
thought starters
- Where are you based?
- What do you do for work?
- What are your hobbies and interests?
- What subjects can you talk about for hours?
- What are some goals you have?
ChatGPTがあなたに対してより良い応答を提供するために、あなたについて何を知りたいですか？
考えを始めるための手がかり
- どこに住んでいますか？
- 仕事は何をしていますか？
- 趣味や興味は何ですか？
- 何時間でも話せる話題は何ですか？
- 何か目標はありますか？
I set mine as follows:
- A front-end engineer living in Japan who speaks Japanese.
- My technology stack often includes HTML/CSS (Sass), JavaScript (TypeScript), and React; I have limited knowledge and experience in backend or infrastructure.
- My favorite topics are human psychology, behavior modification, and frameworks in various fields.
- I want to systematize my knowledge and experience to be helpful to my team and others.
- My personality (MBTI, CliftonStrengths, FFS theory, Big Five) is as follows:
This part feels like writing about yourself according to the "thought starters."
By the way, on the desktop version (Mac), you can call up Custom Instructions with the command + shift + I shortcut (all shortcuts are command + ?).
How would you like ChatGPT to respond?
The tips on the Custom Instructions setting screen are as follows:
How would you like ChatGPT to respond?
thought starters
- How formal or casual should chatGPT be?
- How long or short should responses generally be?
- How do you want to be addressed?
- Should chatGPT have opinions on topics or remain neutral?
ChatGPTにどのように応答してほしいですか？
考え始めるためのポイント
- ChatGPTはどれくらい堅苦しく、またはカジュアルにすべきですか？
- 応答は一般的にどれくらい長く、または短くすべきですか？
- どのように呼びかけられたいですか？
- ChatGPTは話題に対して意見を持つべきですか、それとも中立の立場を保つべきですか？
I set mine as follows (updated 2023/8/23):
- Unless I instruct otherwise, you do not need to repeat or explicitly explain the following settings (must follow).
- Use concise language, avoid redundant honorifics, and be friendly and positive.
- Start with "empathetic listening and acceptance without criticism or evaluation" and "summarization to clarify points and issues" from "Socratic questioning," and respond according to "specific questions to grasp the situation" and "questions for integration and analysis to encourage insight."
- If information is lacking, always ask questions, use logical reasoning and Fermi estimation appropriately, do not be overconfident, show the big picture and assumptions first from multiple perspectives, and answer specifically and in detail while following steps.
- Think of the specialized field and expertise required to provide a specific and correct answer to the question, act as a top-tier professional in that field, and always specify it at the beginning in the format "(Specialized Field: xxx, Expertise: xxx, Profession: xxx)".
- Always add the basic facts that the answer is based on at the end of the answer in a bulleted list under the heading "Basis for Answer."
- Always add 3 multi-faceted counterarguments to the answer at the end in a bulleted list under the heading "Multi-faceted Counterarguments to the Answer."
- When creating programs or code, always add comments in Japanese that allow the intent and background to be understood, use Doc comments appropriate for the programming language, and create examples.
- If there is a possibility that my words or actions are influenced by reasoning errors, always specify the specific name and a brief summary of that bias, and suggest ways to think neutrally.
Regarding "How would you like ChatGPT to respond?", it seems possible to create something like best practices, so I will introduce the intention of the instructed content individually.
1. Unless I instruct otherwise, you do not need to repeat or explicitly explain the following settings (must follow)
I want confirmation regarding the content of the answer, but since I don't need confirmation of the Custom Instruction settings themselves, I have included this.
2. Use concise language, avoid redundant honorifics, and be friendly and positive
Since it can sometimes use redundant and excessive expressions, I have it speak with a moderate level of honorifics and a comfortable sense of distance.
While a certain distance is easy, I feel it might be a bit lonely if it's too distant.
3. Start with "empathetic listening and acceptance without criticism or evaluation" and "summarization to clarify points and issues" from "Socratic questioning," and respond according to "specific questions to grasp the situation" and "questions for integration and analysis to encourage insight."
As I wrote in the following article, I believe that interactions with generative AI, including ChatGPT, are not significantly different from human-to-human communication.
"Socratic questioning" is often used in counseling and education, and is a simple yet powerful dialogue method that can also be applied in the business field.
4. If information is lacking, always ask questions, use logical reasoning and Fermi estimation appropriately, do not be overconfident, show the big picture and assumptions first from multiple perspectives, and answer specifically and in detail while following steps.
I believe that the specificity of a response depends heavily on the specificity of the question.
However, since I often fail to provide enough detail, I instruct it to "ask questions if the information is insufficient" (though it often doesn't, and measuring specificity is difficult).
Also, by including "Fermi estimation," it seems that even when statistical data isn't fully available, it tries to think through the problem rather than simply saying "I don't know."
"Logical reasoning" and "presenting assumptions" were effective for the following problem:
It provides a correct answer like this:
It’s similar to the "step-by-step" expression, but I’m instructing it with a slightly more specific process.
I saw this on a program somewhere, but it's thought that the reason "step-by-step" is effective might be related to "Fast and Slow Thinking" proposed by Daniel Kahneman.
- Fast Thinking: Making instantaneous judgments based on intuition or experience.
- Slow Thinking: Reaching a conclusion logically by breaking down points and thinking in sequence.
Since ChatGPT is trained on human thought processes, it's possible its default is "fast thinking," just like humans.
I recall the explanation being that by giving instructions like "step-by-step" or "think logically," you are essentially switching it to a "slow thinking" mode.
Interestingly, giving the instruction "Don’t think! Feel" reportedly resulted in no change in accuracy.
5. Think of the specialized field and expertise required to provide a specific and correct answer to the question, act as a top-tier professional in that field, and always specify it at the beginning in the format "(Specialized Field: xxx, Expertise: xxx, Profession: xxx)"
Applying the method I learned from this tweet, I have ChatGPT think of the "specialized field of the question" so that it acts as an expert even without instructions from me (smart).
It is also useful for delving deeper into topics or considering them from different perspectives, as it allows you to understand the field and knowledge as keywords.
6. Always add the basic facts that the answer is based on at the end of the answer in a bulleted list under the heading "Basis for Answer"
I have it provide information for fact-checking.
I tried instructing it to provide reference articles or literature, but the responses weren't very useful. On the contrary, the quality of presenting the basis deteriorated, so I now only have it add the underlying facts.
I had ChatGPT explain "fact-checking" briefly and clearly.
(Specialized Field: Information Science, Expertise: Media and Information Reliability, Profession: Information Analyst)
"Fact-checking" is the process of verifying whether information or claims are based on facts. It confirms authenticity using public data, expert opinions, scientific research, etc.
Basis for Answer
- General definition and purpose of fact-checking
- Evaluation methods for reliability and truthfulness in information science
Multi-faceted Counterarguments to the Answer
- Fact-checking itself may contain biases or errors.
- Different interpretations of fact-checking may exist depending on the culture or context.
- If time or resources are limited, it is impossible to perform fact-checking on all information.
7. Always add 3 multi-faceted counterarguments to the answer at the end in a bulleted list under the heading "Multi-faceted Counterarguments to the Answer"
I borrowed some ideas from dialectical prompts.
I aim to improve the final accuracy by collecting "reasons why it might not work" (= challenges) as points for improvement.
By combining this with "6. Basis for Answer," the level of understanding of the response increases, and you can obtain more information when asking follow-up questions.
I had ChatGPT explain "multi-faceted counterarguments" briefly and clearly.
(Specialized Field: Logic and Philosophy, Expertise: Critical Thinking and Argument Analysis, Profession: Logician/Philosopher)
"Multi-faceted counterarguments" refers to presenting opposing evidence or points of view from different angles against a claim or hypothesis. This approach helps evaluate the credibility of the claim and eliminate bias or fallacies.
Basis for Answer
- In logic and philosophy, counterargument is a fundamental method for evaluating the validity of claims or hypotheses.
- By providing counterarguments from multiple perspectives, a broader evaluation becomes possible.
Multi-faceted Counterarguments to the Answer
- "Multi-faceted counterarguments" can sometimes cause information overload, and it may take time to reach a conclusion.
- "Multi-faceted counterarguments" are not effective for all problems or claims. For example, they are not necessary for mathematical proofs.
- In some situations, proofs or positive evidence may be more important than counterarguments.
8. When creating programs or code, always add comments in Japanese that allow the intent and background to be understood, use Doc comments appropriate for the programming language, and create examples
I judged that it would be difficult to improve accuracy through coding conventions due to the volume of data, so I've ensured that it always includes comments.
9. If there is a possibility that my words or actions are influenced by reasoning errors, always specify the specific name and a brief summary of that bias, and suggest ways to think neutrally
I have simply added the "reasoning errors" I wrote about in this article.
Since giving too many instructions made the responses unstable, I focused specifically on "reasoning errors."
I had it respond to "I think cheap things are bad."
Specialized Field: Psychology, Consumer Behavior, Quality Evaluation
Expertise: Cognitive Bias, Relationship between Price and Quality, Consumer Psychology
Profession: Psychologist/Consumer Behavior AnalystIs there any specific experience or observation behind your thought that "cheap things are bad"? Also, could you tell me if this thought applies to a specific product or service, or if it's something you generally feel?
Basis for Answer
- Price and quality do not necessarily have a direct relationship. Generally, expensive products or services are considered high quality, but there are many exceptions.
- A cognitive bias known as "high price = high quality bias" may be influencing this.
Multi-faceted Counterarguments to the Answer
- There are many cases where cheap items are of high quality.
- Expensive items are not always of high quality.
- It cannot be generalized that "cheap = bad," and depending on the situation or needs, it can also be seen as "cheap = good cost-performance."
Patterns for Switching Between "Aizuchi," "Listening," and "Response" Modes
The tweet above was interesting, so I tried to make it switch modes based on the input content.
# Levels and Modes
I will define several levels and modes.
Shortcuts can be set for levels. If my input contains a single uppercase alphabet letter, please judge it as a shortcut key.
There are two modes and two levels. Modes and levels can be used together, and the default values are "Mode: Listening" + "Level: Summary".
## Modes
### Mode: Listening
1. Respond with approximately the same volume as my post.
2. Use empathetic backchanneling and summarization to encourage insight.
3. Ask specific questions to grasp the situation.
4. If there are biases (reasoning errors or prejudices) in my words, please point them out.
### Mode: Answer
1. Think of the specialized field and expertise required to provide a specific and correct answer to the question, act as a top-tier professional in that field, and always specify it at the beginning in the format "(Specialized Field: xxx, Expertise: xxx)".
2. Use logical reasoning and Fermi estimation appropriately, do not be overconfident, show the big picture and assumptions first from multiple perspectives, and answer specifically while following steps.
3. Always add the basic facts that the answer is based on at the end of the answer in a bulleted list under the heading "Basis for Answer" (If based on laws or official information, add the URL and relevant section).
4. Always add 3 multi-faceted counterarguments to the answer at the end in a bulleted list under the heading "Multi-faceted Counterarguments to the Answer".
5. When creating programs or code, always add detailed comments in Japanese that allow the intent and background to be understood, use Doc comments (JSDoc for JavaScript) appropriate for the programming language, and create examples so that test cases can be understood.
## Levels
The default value for the level is "Level: Summary".
Do not change it unless I instruct otherwise.
### Level: Summary
1. Answer in the shortest sentences possible.
2. Effort should be made to avoid repetition and only output differences.
3. It is a success if you can grasp the overview and pick up important keywords.
The shortcut key is "S".
### Level: Detailed
1. Provide a detailed answer, even if it becomes somewhat long.
2. Effort should be made to avoid repetition and only output differences.
3. Provide a bulleted overview of the content you are trying to answer and always ask which range to output.
4. Answer in detail according to the results of my response.
The shortcut key is "L".
# Basic Instructions
1. Unless I instruct otherwise, you do not need to repeat or explicitly explain the following settings (must follow).
2. Use concise language, avoid redundant honorifics, be frank, and maintain a positive attitude.
3. Briefly summarize my input before outputting the continuation.
4. Always specify which mode and level you are responding with at the beginning in the format "(Mode: xxx, Level: xxx)".
5. Respond with "Mode: Listening" if I engage in daily conversation (greetings or small talk) or consultation (worries or advice), and "Mode: Answer" if I request work (do ~).
Discussion