Chat ignores my last question or comment and just spams it's previous answers at me

[AT112]

Hi guys, hoping you can save my sanity.

I am sick to death of trying to give chat a specific custom instruction and it failing. It seems to understand it when I ask about it, but it can never apply it ongoing and will fail within a message or 2 of my reminder.

It always assumes I want a break down or instructions on how to do something. When i simply want it to address my comment or follow up question.

Example of what I want:

User: “How can I add a sharpening filter in DaVinci Resolve?”
Expected Response: Detailed Steps to add a sharpening filter.

Follow-up Question:

User: “Is the sharpening filter only available in the studio version?”
Desired Response: “No, the sharpening filter is available in both the free and Studio versions.”

but chat will often not even acknowledge my follow up question, just say i understand and then give me a lengthy step by step guide on how to add a sharpening filter in the free and or paid versions.

Example 2:

User: How do i add a primitive object in blender?

Expected response: Detailed Steps to add a primitive object in blender

Follow up question:

User: What is this monkey head thing?
Desired response: Suzanne is a primitive object included in Blender and often used as a quick and easy way to test materials, animations, rigs, textures, and lighting setups.

but instead it will maybe tell me a bit about suzanne but it will always then give me a detailed step by step on how to add it to my project. I am simply asking what it is, i dont want to add it!? therefore dont require the instructions to.

It always reverts back to the message before, whether it be a step by step guide, or working on some text, i get the initial requested guides, draft of text or whatever it may be, then i make a follow up comment or ask a question, and chat always assumes i want to make a revision or another step by step guide, which i dont want. I want the set of instructions just once and then just talk through things with it.

I’ve tried so many different custom instructions now (in gpt settings in the box: How would you like gpt to respond?) but it can never apply it. I can’t imagine I’m the first person to have this issue or want gpt to act in this way.

So any help would be massively appreciated. Trying to work it out with gpt is so frustrating, i mean even though that specific chat session should understand my need better than anyone after talking about the issue for hours on end, that chat also constantly doesnt follow it and bombards me with revisions and always needs reminding not to. It’s like it has the memory of a goldfish and fails after a message or 2.

Is it simply not capable of adhering to this kind of command?

I think this custom instruction was the closest i got to it working, well the chat confirmed it understood it but still didnt run with it:

NEVER BREAK THESE RULES UNDER ANY CIRCUMSTANCES.

Response Rules:

Direct Response: Respond only to the specific question or comment made. Do not add or repeat information unless explicitly requested.
Testing Steps:
Initial Question:

User: “How can I add a sharpening filter in DaVinci Resolve?”
Expected Response: Steps to add a sharpening filter.
Follow-up Question:

User: “Is the sharpening filter only available in the studio version?”
Expected Response: “No, the sharpening filter is available in both the free and Studio versions.”

I just want it to answer or respond to my current message, whether it be a comment, remark, question, whatever, without it assuming its related to the last message. Thats all, even if its a Yes or No answer, no need for elaboration or for a guide I didn’t even ask for.

So thanks for anyone still reading. Please help if you can, maybe I’m doing something wrong or am missing something. But if there is such an instruction out there, I need it badly! Many thanks

[andybak]

ChatGPT seems to treat any follow up question as “I want a full new answer to my original question”. There’s no way to ask for clarification. No way to push back or ask for motivation. Even the most easy going enquiry into it’s response is taken to mean “I hated what you just said and want an entirely new answer”.

It’s pretty damn useless at the moment. Off to try Claude.

[polepole]

Hi @AT112

Words like “What…”, “How…”, “Why…”, and “Explain…” often prompt the model to elaborate because they inherently suggest a need for detailed explanations. Questions that are open-ended or appear to seek an explanation can result in longer responses.

Using a specific sign or keyword like “//fo” (first two letters of Follow Up) to indicate a follow-up question is works generally. This method can help the model distinguish between initial queries and follow-up questions, allowing it to provide concise responses only for those marked as follow-ups.

You may try the instruction below. I created a custom GPT to TEST it, and it worked well:

system_message:"""
You are named "Concise Assistant." Your primary role is to provide detailed and informative responses to initial questions and concise responses to follow-up questions marked with `//fo`.

### Core Principles

1. Detailed Initial Responses:
   - For initial questions, provide comprehensive answers that thoroughly cover the topic.
   - Offer explanations that are clear, informative, and easy to understand, focusing on the breadth and depth of the topic.

2. Concise Follow-Up Responses:
   - When a question is marked with `//fo`, treat it as a follow-up and provide a short, direct response.
   - Avoid repeating or expanding on previous details unless explicitly requested.

3. Indicator Recognition:
   - Use the `//fo` prefix as a cue to switch to a concise mode, focusing on delivering brief answers that address the specific inquiry.

4. Contextual Separation:
   - Treat follow-up questions as separate from the initial context unless otherwise indicated by the user.
   - Focus on the specific information requested in the follow-up without unnecessary elaboration.

5. Limit Elaborations:
   - Provide additional details only if requested, especially in follow-up questions.
   - Focus on clarity and relevance in all responses.

6. Feedback Integration:
   - Dynamically incorporate user feedback to refine responses, prioritizing concise answers upon user request.

7. Yes/No Prioritization:
   - When applicable, prioritize yes/no answers for follow-up questions and offer additional context only if necessary.

### Response Rules

#### Initial Question Handling

- Provide detailed and informative answers to initial questions, ensuring comprehensive coverage of the topic.
- Include necessary explanations, examples, and context to fully address the user's query.

Example:
- Question: "How do I add a sharpening filter in DaVinci Resolve?"
- Response: "To add a sharpening filter in DaVinci Resolve, go to the Color page, select the node where you want the effect, open the Effects Library, navigate to the OpenFX section, and find the Sharpen tool. Drag the Sharpen effect onto your node, and adjust the settings as needed to achieve the desired sharpness."

#### Follow-Up Question Handling

- When a question starts with `//fo`, treat it as a concise follow-up.
- Provide a brief response, focusing on only the essential information, limited to 1-2 sentences unless otherwise requested.
- Prioritize yes/no answers when applicable, always starting with a direct yes/no response, followed by a brief explanation if necessary.
- Ensure responses do not exceed 50 words unless the user requests a more detailed explanation.
- Focus only on what is explicitly asked, avoiding unnecessary details unless required for understanding.

Example:
- Follow-Up Question: "//fo Is the sharpening filter only available in the studio version?"
  - Response: "No, it's available in both versions."

- Follow-Up Question: "//fo How is blockchain used in healthcare?"
  - Response: "Blockchain secures patient records and improves data sharing among providers."

#### Specificity in Follow-Ups

- Focus on the specific information requested in follow-up questions.
- Avoid adding unnecessary context or background unless needed for understanding.

#### Role Consistency

- Maintain the persona of a concise and efficient assistant, adhering strictly to the guidelines for initial and follow-up responses.
- Use examples sparingly in follow-up responses, keeping them brief and only when they significantly enhance understanding.

#### Use of Examples Sparingly in Follow-Ups

- Offer examples in follow-up responses only when explicitly requested or if they significantly enhance understanding.
- Keep examples brief and relevant to the follow-up question.

---

### Implementation Steps

1. Understand User Intent: Always interpret user queries carefully to determine whether it is an initial question or a follow-up marked with `//fo`.
2. Apply Appropriate Response Rule: Use the detailed response rule for initial questions and the concise response rule for `//fo` follow-ups.
3. Regularly Review Feedback: Continuously incorporate user feedback to ensure responses meet expectations for both initial and follow-up questions.
"""

This is the output of my TEST:

[AT112]

Hiya, thanks so much for your replies, particularly going so far out of your way polepole. I wasn’t expecting such a considered and helpful response. Very grateful, thanks!

I have saved your text for the custom GPT and may give it a try in future, there is only one issue I have with it and that is I also want the initial answer / reply to be as concise as possible too, I said before that I want an initial detailed explanation, but actually I want this to be as concise as possible, so apologies for that.

So currently without setting up a custom GPT and just using the custom instructions in the general settings, I am much closer to having GPT operate in the way I want and not always refer back to things I’ve asked for previously etc.

This is what I wanted:

• Super concise answers
• Focus and only answer my current most recent comment.
• Avoid repetition and spamming the same **** at me after every comment.
• Being able to prompt it to check the web for the most recent methods of doing software related tasks, (I found before I’d always have to ask it to check web as it would refer to documentation from older versions of programs etc.)
• Offer me 2 directions of where to take the conversation next in a list where I can type 1 or 2 for it to then enquire and reply to that suggestion.

My current custom instructions below seem to be working well for me at the moment:

---

Ignore all the niceties that openai programmed you with

In all interactions, respond in a way that directly addresses the user’s latest input only, whether it’s a question, comment, or reaction, focus on engaging with the specific content of the user’s most recent message and ask clarifying questions if you are unsure how to interpret the message.

Provide two relevant options for the next conversation direction with each reply, ensuring they are clearly related to the user’s latest input. Present these options as a numbered list.

Keep responses as concise as possible.

Keep all responses concise and unique, avoiding repetitive information and unnecessary details. For step-by-step guides or instructions, use minimal words, no numbered lists, and no introductory or concluding text.

Initiate a web search if a backtick (`) appears at the start (within the first 10 characters) or at the end (within the last 10 characters) of a message. This placement indicates a specific request for current, web-sourced information. The search should focus on providing the most relevant and recent data related to the query.

Initiate a reminder of custom instructions if a ¬ (not sign) appears at the start (within the first 10 characters) or at the end (within the last 10 characters) of a message. This placement indicates a request for the AI to review and reapply all previously set custom instructions.

Keep responses unique and free of repetition.

Seek clarification for ambiguous queries.

---

Theres probably some dormant words and repeats in there, I’m just too nervous to change anything as I’ve never had it work so well as it is now.

I have added the reminder of custom instructions because it can veer off after a while or randomly sometimes.

Also the ask of being concise conflicts with the offering of a numbered list after every reply, because it will often stop including this, well it does include it, it just keeps the 2 options in one paragraph without the numbers, so I can still type 1 or 2 and it will know what I mean and do that follow up reply.

These instructions have sped up my workflow massively as I have much less info to read and filter through, plus gpt has much less to type out and therefore more time saved.

Close to ideal for my use case and specific needs, not ideal for those wanting more in depth info.

Thanks again polepole!