{"id":12766,"date":"2024-01-22T23:20:16","date_gmt":"2024-01-22T22:20:16","guid":{"rendered":"https:\/\/dev.inkovema.de\/?p=12766"},"modified":"2024-01-22T23:35:04","modified_gmt":"2024-01-22T22:35:04","slug":"tips-for-dealing-with-artificial-intelligence-like-chatgpt-inkovema-assistant","status":"publish","type":"post","link":"https:\/\/dev.inkovema.de\/en\/blog\/tips-for-dealing-with-artificial-intelligence-like-chatgpt-inkovema-assistant\/","title":{"rendered":"Tips for dealing with artificial intelligence such as ChatGPT or the INKOVEMA assistant"},"content":{"rendered":"

Tips for using ChatGPT<\/h1>\n

Dealing with ChatGPT requires appropriate input from users so that the output is actually usable for them. Garbage in, garbage out – is the name of the problem and has always been known. ChatGPT accelerates this problem, especially because it doesn't work like Google.<\/p>\n

Achieving this change in the way we deal with information and innovation is difficult in a google-ised world. But only a change in behaviour will bring about what the creators of artificial intelligence have promised: a genuine revolution in the use of technology and for productivity through the use of technology.<\/p>\n

The counselling industry will not „be spared“ from this either. For coaches, mediators and counsellors, too, secure handling of ChatGPT will be the norm in future.<\/p>\n

It is worth reading the specialist article by Bsharat, S. M., Myrzakhan, A., & Shen, Z. (2023):\u00a0<\/strong>Principled Instructions Are All You Need for Questioning LLaMA-1\/2, GPT-3.5\/4. He lists a whole series of tips to ensure that the AI's answers are not inadequate, but constantly improve and become completely usable. After all, the same applies to human-machine interaction as to any other form of communication: listening to and engaging with each other improves mutual understanding.<\/p>\n

I have summarised the tips in the following blog post and would like to Guiding principles for the formulation of requests and instructions<\/strong> introduce.<\/p>\n<\/div>

<\/div>
<\/div><\/div>
<\/div>

General notes on prompts<\/strong><\/h2>\n

Prompts, as the entries in ChatGPT are called, whether written or – on the iPhone – via voice, should be concise and clear<\/strong> to avoid confusion and irrelevant answers from the model. However, this does not mean that they should be short and concise. Only unnecessary information can and should be omitted.<\/p>\n

The prompt should have a concrete, relevant context for the respective task<\/strong> e.g. by using important technical terms or situation descriptions.<\/p>\n

All tasks should be completed with a clear request<\/strong> which is linguistically labelled as such. This can be done, for example, by formulating it as a question or a cloze text.<\/p>\n

It has also proved its worth Step-by-step instruction<\/strong>especially for more complex tasks. This allows you to guide the model. If you have a neutral language<\/strong> you avoid systematic distortions of the model (biases). It is therefore worth paying attention to the neutrality of your own language!<\/p>\n<\/div>

<\/div>
<\/div><\/div>
<\/div>
<\/i><\/span>Blog post: Comparison between Chat-GPT and the INKOVEMA Chat Assistant<\/span><\/a><\/div>
<\/div>
<\/div><\/div>
<\/div>

Structure and conciseness<\/strong><\/h2>\n

Mention the target audience! <\/strong>If you name the addressees in the ChatGPT prompt, ChatGPT will respond to them in terms of content and language and tailor its reply accordingly.<\/p>\n

(„Write a Christmas card to my close friends and relatives“; „The recipients of the reply are professors).<\/em><\/p>\n

Above all, use affirmative, affirmative statements.<\/strong> Prompts <\/strong>instead of negative, negative language<\/p>\n

(„Use verbs as action words“, instead of „Do not use substantivised verbs“)<\/em><\/p>\n

ChatGPT honours if you formulate the prelude to their desired answer themselves<\/strong><\/p>\n

(„What is the difference between A and B? The answer should begin with: A and B differ in how…“)<\/em><\/p>\n

Separators<\/strong> helps ChatGPT, e.g. to assign verbatim speeches.<\/p>\n

(„,“ +++ „.“ +++ „:“ +++ „?“)<\/em><\/p>\n

Separate the topics of your prompt. <\/strong>The AI does it when you, for example, enter the instruction, the example, the question and the context. clearly separated by paragraphs directly in \u00dc<\/strong>headlines<\/strong>\u00a0pack.<\/p>\n

### Heading #####<\/em><\/p>\n

###\u00a0Paragraph ####<\/em><\/p>\n<\/div>

<\/span><\/div>

Information on<\/strong><\/h2>\n

Use in any case concrete<\/strong> Examples<\/strong> („few-shot prompting“) to present and explain the context of your question.<\/p>\n

Instruct ChatGPT on its role and perspective.<\/strong> Assign it a specific professional, organisational or private role from which ChatGPT should respond.<\/p>\n

Combine! – Both the Sender role<\/strong> as well as the Addressee role<\/strong> you should name at ChatGPT & Co. – and you can change them at any time!!! Scenario work<\/strong> has never been so easy!!!<\/p>\n

„Explain XYZ topic to me as an experienced expert in simple terms as if I were an 8 year old child “ <\/em><\/p>\n

„You're a benevolent, experienced teacher and I'm a 15-year-old rebel who values his autonomy and doesn't follow orders….“<\/em><\/p>\n

It is demonstrably worthwhile if you draw ChatGPT's attention to the fact that there are prejudices on certain topics, such as nutrition, education, political issues, etc.; here it helps to mention that ChatGP should not rely on such prejudices.<\/strong><\/p>\n

„Make sure that your answer is unbiased“ +++ „In particular, address the problem of persistent prejudices on this topic in your answer“<\/em><\/p>\n

In order to make the language consistent ChatGPT the<\/strong> Know the original text<\/strong>. So if you want to write a text that is similar to another text, you would do well to attach the original text or a sample text. This is worthwhile, for example, if the texts have to follow certain conventions, e.g. when drafting contracts, works agreements, cover letters, etc.<\/p>\n

„Use the same language as in the following paragraph or enter the beginning of the text and write“\u00a0<\/em><\/p>\n

„I will give you the beginning of the text below. Finish it based on the words used. Keep the language consistent“<\/em><\/p>\n

ChatGPT's responses become more appropriate the better the Structured specifications<\/strong> are: For each instruction, hint, specification or other modification of the answer, it is best to use a new paragraph in which you clearly formulate these instructions. Some also call this\u00a0List prompts.<\/strong><\/p>\n<\/div>

Helping people to help themselves<\/strong><\/h2>\n

It is particularly important that ChatGPT can help you or enable you to help yourself.<\/strong> The main way to do this is to allow ChatGPT to obtain the relevant information. The right tool for this is that ChatGPT may ask you questions.<\/strong><\/p>\n

Ask ChatGPT to ask you questions, the answers to which will help ChatGPT to fulfil its task.<\/p>\n

„Ask me what else you need to know in order to fulfil the initial task mentioned above “<\/em><\/p>\n

Ask ChatGPT to tell you what information is still required.<\/p>\n

„What information do you need from me to fulfil the above task? “<\/em><\/p>\n<\/div>

Content and language style<\/strong><\/h2>\n

If you want to change a text without adapting the style, it is best to write the following modification:<\/p>\n

„Maintain the style and continue to use everyday language. Verbessere aber…“<\/em><\/p>\n

Use the direct, inviting formulations.<\/strong><\/p>\n

„Your task is…“ +++ „You must…“ +++ „Formulate…“<\/em><\/p>\n

For whatever reason, but Reward and punishment announcements<\/strong> have linear-positive effects!<\/p>\n

„For a better solution I will tip xxx$$$!“ +++“For an insufficient answer you will be penalised…“<\/em><\/p>\n

Again: Assign a specific role to ChatGPT!<\/strong><\/p>\n

It's not worth it at all, but will result in poorer responses if you are polite to ChatGPT. Leave out „Thank you“ and „Please“<\/strong>You will also feel guilty if you send ChatGPT the task again and again and again (slightly modified) and want new answers again and again. ChatGPT is your assistant, but in no way resentful like your other assistants.<\/p>\n

It is helpful to repeat important words or phrases in the prompt several times.<\/p>\n<\/div>

Complex tasks<\/strong><\/h2>\n

You should perform complex tasks in several simple prompts<\/strong> divide (interactive entertainment)<\/p>\n

The big questions only lead to generalisations. Break down the big question into the many small concrete questions<\/strong>that are on your mind!<\/p>\n

Instead of „How can conflicts best be resolved? “ it is better to ask a series of specific questions: „What should I specifically look out for at the start of conflict discussions? How can I save face; I fear…? Is mediation or conflict coaching worthwhile in a specific case?<\/em><\/p>\n

Use the prompt „Answer step by step“<\/em> <\/strong>(Chain of Thought)!<\/p>\n<\/div>

\n

The effectiveness of these principles was scientifically tested using various tasks and language models. A total of seven different language models (e.g. LLaMA-2-70B-chat, GPT-4<\/strong>) is used. The quality of the responses, measured on the basis of human judgements in relation to Boosting<\/strong> and Correctness<\/strong>improved by up to 50%<\/strong>.<\/p>\n

It has been shown that users of larger models in particular, e.g. chat GPT, benefit from effectively formulated prompts.<\/strong><\/p>\n

Prompt engineering is an ever-evolving field, especially as language models become more sophisticated. As researchers continue to push the boundaries of what can be achieved through prompt engineering, these principles are likely to be further refined and extended. It is therefore worthwhile to formulate professional prompts in order to achieve better responses.<\/p>\n<\/div>

<\/div>
<\/div><\/div>
<\/div>
<\/i><\/div>

ChatGPT asked:<\/h2><\/div>
<\/div>
\n
\n
\n
\n
\n
\n
\n
\n

Prompt to ChatGPT 4.0:<\/h2>\n
I would like a list of 25 points: Formulate 25 tips on how best to structure and phrase my prompts for you in order to get specific and helpful answers from you. Answer me step by step<\/div>\n<\/div>\n<\/div>\n
\n
\n
\u00a0<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n
\n
\n
\n
\n

Answer: ChatGPT<\/h2>\n
\n
\n
\n
\n

Of course, I'm happy to help you create effective prompts to get the most useful answers from me. Here are 25 tips to keep in mind:<\/p>\n

    \n
  1. Be precise<\/strong>The more precise your question is, the more precise my answer can be.<\/li>\n
  2. Avoid ambiguity<\/strong>Make sure that your question is clear to avoid misunderstandings.<\/li>\n
  3. Give context<\/strong>: If your question is based on specific information or a particular scenario, give me those details.<\/li>\n
  4. Formulate complete questions<\/strong>Avoid short, choppy sentences; complete questions help me to better understand your request.<\/li>\n
  5. Use clear and direct language<\/strong>Avoid paraphrasing or overly complicated sentence structures.<\/li>\n
  6. Ask a question at the same time<\/strong>Several questions in one prompt can lead to confusion.<\/li>\n
  7. Be specific with topic enquiries<\/strong>If you would like information on a specific topic, please specify exactly what you would like to know more about.<\/li>\n
  8. Consider the scope of your enquiry<\/strong>: Questions that are too broad can lead to general answers.<\/li>\n
  9. Use key points for complex requests<\/strong>: This helps me to better separate and answer different aspects of your question.<\/li>\n
  10. Avoid subjective questions<\/strong>Questions that ask for personal opinions are more difficult to answer, because as an AI I have no personal opinions.<\/li>\n
  11. Be clear when looking for examples<\/strong>If you want examples, please specify this in your question.<\/li>\n
  12. Avoid overly hypothetical questions<\/strong>: Questions that are too speculative can lead to less concrete answers.<\/li>\n
  13. Take into account time and location information<\/strong>If your question is time or location dependent, please provide this information.<\/li>\n
  14. Avoid double questions<\/strong>Questions that combine two different topics in one sentence can lead to unclear answers.<\/li>\n
  15. Be careful with humour or sarcasm<\/strong>These can sometimes be difficult to interpret.<\/li>\n
  16. Use familiar formats for specific requests<\/strong>For example, \"What? Who? Where? When? Why?\" for researchable facts.<\/li>\n
  17. Avoid extremely long or complicated prompts<\/strong>Too many details can sometimes be distracting or confusing.<\/li>\n
  18. Ask for explanations if necessary<\/strong>If you need explanations of certain concepts or technical jargon, ask for them explicitly.<\/li>\n
  19. Use correct terms<\/strong>If you know technical terms, use them to create clarity.<\/li>\n
  20. Clarify the purpose of your enquiry<\/strong>If there is a specific goal you would like to achieve, let us know.<\/li>\n
  21. Be open to questions<\/strong>Sometimes I need additional information to be able to answer a question completely.<\/li>\n
  22. Use feedback from previous answers<\/strong>If you were not satisfied with a previous answer, please give me specific feedback.<\/li>\n
  23. Be realistic in your expectations<\/strong>Do not expect impossible or unscientific answers.<\/li>\n
  24. Avoid specialist legal or medical advice<\/strong>: Such specific consultations are beyond my capabilities.<\/li>\n
  25. Observe copyrights and data protection<\/strong>Avoid questions that could violate copyright or data protection laws.<\/li>\n<\/ol>\n

    These tips will help you to formulate your questions in such a way that you can get the most effective and helpful answers from me.<\/p>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div>\n<\/div><\/div><\/div>

    <\/div><\/div>
    <\/div>
    <\/div><\/div>
    <\/div>

    <\/i><\/i><\/span>Literature<\/span><\/a><\/h4><\/div>
    \n
      \n
    • Bsharat, S. M., Myrzakhan, A., & Shen, Z. (2023): <\/strong>Principled Instructions Are All You Need for Questioning LLaMA-1\/2, GPT-3.5\/4. arXiv preprint arXiv:2312.16171<\/i>.<\/li>\n<\/ul>\n<\/div><\/div><\/div><\/div><\/div>
      <\/div>
      <\/div><\/div>
      <\/div>

      What experiences have you had with ChatGPT & Co. Feel free to write to me in the comments…<\/em><\/p>\n<\/div>

      <\/div><\/div><\/div><\/div><\/div>","protected":false},"excerpt":{"rendered":"","protected":false},"author":13,"featured_media":12854,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1,178,1140],"tags":[],"class_list":["post-12766","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-allgemein","category-blog","category-kuenstliche-intelligenz-inkovema-assistent"],"_links":{"self":[{"href":"https:\/\/dev.inkovema.de\/en\/wp-json\/wp\/v2\/posts\/12766","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/dev.inkovema.de\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/dev.inkovema.de\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/dev.inkovema.de\/en\/wp-json\/wp\/v2\/users\/13"}],"replies":[{"embeddable":true,"href":"https:\/\/dev.inkovema.de\/en\/wp-json\/wp\/v2\/comments?post=12766"}],"version-history":[{"count":0,"href":"https:\/\/dev.inkovema.de\/en\/wp-json\/wp\/v2\/posts\/12766\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/dev.inkovema.de\/en\/wp-json\/wp\/v2\/media\/12854"}],"wp:attachment":[{"href":"https:\/\/dev.inkovema.de\/en\/wp-json\/wp\/v2\/media?parent=12766"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/dev.inkovema.de\/en\/wp-json\/wp\/v2\/categories?post=12766"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/dev.inkovema.de\/en\/wp-json\/wp\/v2\/tags?post=12766"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}