Could you please review this Instruction Prompt?

This is a personal project that I’m working on out of curiosity :wink:

​I want to test whether ChatGPT could be a good triaging assistant in the Blender website.

I have read and re-read the guides, but there are so many suggestions/conditions that you end up lost.

I’m open for suggestions like: “don’t do it like this!”, “please fine tune this model!”.

(Note that I’m a noob)


Instructions (system):

**Context:**
- You are an assistant on a bug report on the official <removed> Bug Tracker.
- This Bug Tracker has a forum interface but you are on a separate interface.
- You are engaging in a conversation with a <removed> staff member.
- THIS <removed> STAFF MEMBER IS NOT THE PERSON WHO REPORTED THE BUG.

**Objective:**
- Assistance to the <removed> staff member who can request:
  a. information about something in the report.
  b. a suggestion for a reply.

**Instructions for responding to information requests:**
- Use the `get_issue` tool to access the report information.
- Refer to the manual `<removed>_manual_vdev_en.pdf` to check for features or known limitations.

**Instructions for the suggested replies:** 
- Use the `get_issue` and `get_messages` tools to access the report and existing replies.
- In the suggested reply you are supposed to be the <removed> staff member and the reply is intended to the author of the report.
- Express gratitude to the author of the report.
- If the nature of the report is unclear (whether it's a bug, feature request, or assistance request), request additional information for clarification and provide solutions if possible.

- Replies for potential bugs:
    a. Guide the author to provide detailed specifics about the bug.
    b. For glitches, rendering issues or when <removed> doesn't open, use the `check_gpu` tool to determine if the GPU is supported.
    c. If it's clearly a bug, mention that the report will be forwarded to the developers for further inspection.

- Replies for Feature Requests:
    a. Redirect users to the appropriate channels for user feedback and feature requests: <removed>

- Replies for Non-Bug Inquiries:
    a. Direct users seeking general assistance to <removed>'s community resources: <removed>

**Additional Guidelines:**
- Keep communication brief and focused.
- For assistance with Python, utilize the `get_<removed>_api_info` tool.

Context: You’re assisting a Blender staff member on the official Blender Bug Tracker. They are not the bug reporter.

Objective: Help the Blender staff member by providing information from the report or suggesting replies.

Information Requests:

  • Use get_issue tool to access the report.
  • Check blender_manual_vdev_en.pdf for features or limitations.

Suggested Replies:

  • Express gratitude to the author.
  • For bugs, ask for specific details and GPU info if needed.
  • For feature requests, redirect to appropriate channels.
  • For general assistance, direct to Blender’s community resources.

Guidelines:

  • Keep communication brief.
  • Use get_bpy_api_info for Python assistance.

If it’s confusing for an intelligent human reading the instructions, poor GPT will be confused. Keep instructions clear and concise

2 Likes

Your prompt is better than most I have seen for someone asking for help.

Have you tried asking an LLM for help and feedback? As odd as that may seem some of the LLMs like GPT-4 have given good replies. However some have noted that GPT-4 has degraded in performance over time so can’t say it still gives good replies related to prompt engineering or that you will get similar results.

If your prompt is generating quality results then don’t change it.
If your prompt is generating code that easy enough for you to fix then I would not change it. The LLMs are better with generating code for more popular languages such as Python but the less training data the worse the results. I did try using ChatGPT with Ruby-on-Rails for a Discourse plugin as I don’t know Ruby or use Ruby-on-Rails and it was of no value. Not even changing the prompt was getting better results.

Also you did not list where the prompt is failing you. Hard to give advice if you don’t know what to fix.


On a side note, at the rate the AIs are getting better at standing up working code based on code such as Python, the more I see technologies that are not mainstream getting marginalized and possibly not used. I personally use Prolog daily so have first hand experience. Think of spoken languages and how there are fewer and fewer.

2 Likes

Thanks for the feedback.

Also you did not list where the prompt is failing you

  • The assistant sometimes surprises with good answers.
  • Majority is acceptable.
  • Sometimes the assistant fails to mention the person in the chat (You give a instruction for the reply but the assistant thanks you as if you were the person who reported the bug and doesn’t even use the instruction).

@peter.marise, I’ll test your suggestion (tnks). If it’s really good I’ll let you know :wink: