2023-12-29 ChatGPT Prompt Guidlines

1. Most important words are FIRST tokens use only following (2 first words) https://github.com/tatsu-lab/stanford_alpaca#data-generation-process

   

2. System Prompt examples: https://matt-rickard.com/a-list-of-leaked-system-prompts "[INST]\n<>\n{system_prompt}\n<>\n\n{user_prompt}[/INST]"

    

3. User Prompt structure:

   First start with Task (in english) - This can also be “System prompt” so you do not need to add it user prompt at all a) Task: classify sentence …

   b) Context, persona - Act as Steve Jobs in a Role as "Innovation Officer" at Apple preparing a Report to evaluate new product proposals.

   d) Formatting Tone: Mark changes in the Answer in Bold. Table, CSV, formalize as code, HTML, Question, Summary, Tweet 1-2 paragraphs c) Facts: “…..”

    

   

4. Every prompt in new session

5. Task and keywords in English

6. Task starts with verb

7. Terms and Keywords with Capital Case Letters

8. Multiple word facts delimited with quotes, like "Market Research"

9. Fix grammar errors

10. Add as much as possible Facts and Examples

11. Content and answers can be multi-lingual, but prompt must be in English

12. Do not chain together multiple data-mining prompts, each prompt should be in a new session

13. Limit length of content + prompt to maximum of 1500 words (for now up to GPT4). To get longer prompts need to summarize content

14. Seperate code blocks using Markdown like

    3
    1
    ```python
    2
    something = True
    3
    ```
    

15. For generative content limit output to as much as specific rules, like length of answer, style of answer, etc. For example:

    8
    1
    ...
    2
    Write 1-3 sentences in English using academic style based on information above:
    3
    
    4
    ...
    5
    Write 1-3 paragraphs in Latvian using academic style based on information above:
    6
    
    7
    ...
    8
    Summarize in one short sentence using academic style describing it as related work in Latvian:
    

16. Method to improve answers with followup prompts (https://arxiv.org/pdf/2303.17491.pdf)

    3
    1
    1. ....
    2
    2. Find problems with your answer.
    3
    3. Based on this, the improve answer.
    

17. For zero-shot classification use Capital letters as classes instead of lower case letters, try to use “,” instead of “or” / “and” words "Yes", "No"

18. Separators such as triple quotes (""" """), XML tags ( ), or section headings can help the AI separate sections of text that should be treated differently. Especially important when you do not want to include this content as prompt itself, but just as source content.

19. ChatGPT cannot count exactly, but will usually roughly follow any word or character limits given in the prompt. Avoid any numerical estimations or tasks

20. Avoid as much as possible any rare token combinations like people or company names as these would not be present in ChatGPT dataset and content/prompt would get significantly worse results. You should replace rare words with some other word before feeding in LLM.

21. Do not combine multiple tasks into single classification if high precision is required (do not make prompt to combine Is this content in Latvian and in academic style, answer using only: "Yes", "No"), instead for each classification prompt make separate prompt

22. ⚠️Temperature mining. When asking to generate prompt use high temperature of model, generate multiple answers and select answer that is the closest semantically to question. Or use multiple answers to summarize in to a single answer with another prompt.

    7
    1
     completion = openai.ChatCompletion.create(
    2
            model="gpt-3.5-turbo",
    3
            max_tokens=500,  # max output tokens
    4
            temperature=1.0, # will generate more diverse answers
    5
            n=3,
    6
            messages=messages,
    7
        )
    

     

23. Use positive rather than negative phrases, e.g. "write formally" rather than "do not write informally".

24. Be very specific about CONTENT part do not write “text above”, use “question above”, “answer above”, “facts above”, “list of student characteristics above”

25. Do not write words that are not necessary, be as clear and precise as possible: Although communication with ChatGPT sometimes feels almost human, pleasantries like "please" and "thank you" are unnecessary. ChatGPT does not need eloquent instructions, but clear ones.

26. Zero-shot tasks that ChatGPT can do:

    1. Summarizing content

    2. Generating questions, answers, titles, etc. based on content

    3. Paraphrasing content

    4. Translating content to your native language

    5. Classifying content (Sentiment positive, negative or neutral)

    6. Change tone of content (from positive to negative)

    7. Change theme or style of content

Other guidlines

Prompt Engineering for Effective Interaction with ChatGPT - MachineLearningMastery.com

ChatGPT Guide: Use these prompt strategies to maximize your results (the-decoder.com)

https://github.com/dair-ai/Prompt-Engineering-Guide

OpenAI cookbook https://github.com/openai/openai-cookbook

https://github.com/brexhq/prompt-engineering

https://www.promptingguide.ai/

https://prompts.chat/

Lilian Weng https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/

Latest benchmarks

https://gpt4all.io/index.html