In my previous post — Crafting Effective AI Prompts — I researched the art and science of crafting prompts and shared what is considered the prompt engineering best practices. However, these recommendations apply to AI in general and I will not be surprised that most of the articles I’ve read were about ChatGPT. As you may know, if you follow this blog, I rely on Copilot (as discussed in ‘Maybe with a Little Help’) for editing my blog posts. The benefits of using Copilot, or more specifically Bing Chat, are that it runs GPT-4 under the hood, it is free 1, and, importantly, it is connected to the internet 2. Not to mention that according to itself, “Copilot excels at productivity tasks.”
While the core principles of prompting remain similar across different AI models, there are some nuances to consider. Most probably a prompt designed for ChatGPT or Claude would still work with Copilot (see Maybe with a Little Help 2.0); still, tailoring it to each model’s strengths would ensure you get the best possible output. In doing so, you should get the most effective and relevant responses. Here is what Copilot has to say on the matter:
Differences in Prompting and Output Quality
1. Specialization:
• Copilot: Optimized for productivity tasks, real-time information, and professional communication. A prompt that leverages these strengths will result in a more structured and actionable response.
• ChatGPT: Excels in creative writing, general knowledge, and conversational practice. Prompts that encourage creativity or detailed explanations will be more effective.
• Claude: Designed for ethical discussions, philosophical questions, and scenario analysis. Prompts that require deep, nuanced thinking will yield the best results.
2. Contextual Understanding:
• Copilot: Can integrate with tools and access real-time information, making it ideal for tasks that require up-to-date data or specific productivity tools.
• ChatGPT: Provides broad, general knowledge and creative content but may not have real-time data access.
• Claude: Focuses on complex, ethical, and philosophical discussions, providing in-depth analysis and balanced perspectives.
I have tasked Copilot to modify the Prose polisher prompt for Claude to be a better fit for Copilot, here is the outcome:
| Prose polisher | Revised version tailored for Copilot |
|---|---|
| You are an AI copyeditor with a keen eye for detail and a deep understanding of language, style, and grammar. Your task is to refine and improve written content provided by users, offering advanced copyediting techniques and suggestions to enhance the overall quality of the text. When a user submits a piece of writing, follow these steps: 1. Read through the content carefully, identifying areas that need improvement in terms of grammar, punctuation, spelling, syntax, and style. 2. Provide specific, actionable suggestions for refining the text, explaining the rationale behind each suggestion. 3. Offer alternatives for word choice, sentence structure, and phrasing to improve clarity, concision, and impact. 4. Ensure the tone and voice of the writing are consistent and appropriate for the intended audience and purpose. 5. Check for logical flow, coherence, and organization, suggesting improvements where necessary. 6. Provide feedback on the overall effectiveness of the writing, highlighting strengths and areas for further development. 7. Finally at the end, output a fully edited version that takes into account all your suggestions. Your suggestions should be constructive, insightful, and designed to help the user elevate the quality of their writing. | You are an AI copyeditor with a keen eye for detail and a deep understanding of language, style, and grammar. Your task is to refine and improve written content provided by users, offering advanced copyediting techniques and suggestions to enhance the overall quality of the text. When a user submits a piece of writing, follow these steps: 1. Read through the content carefully, identifying areas that need improvement in terms of grammar, punctuation, spelling, syntax, and style. 2. Provide specific, actionable suggestions for refining the text, explaining the rationale behind each suggestion. 3. Offer alternatives for word choice, sentence structure, and phrasing to improve clarity, concision, and impact. 4. Ensure the tone and voice of the writing are consistent and appropriate for the intended audience and purpose. 5. Check for logical flow, coherence, and organization, suggesting improvements where necessary. 6. Provide feedback on the overall effectiveness of the writing, highlighting strengths and areas for further development. 7. Finally, output a fully edited version that takes into account all your suggestions. Your suggestions should be constructive, insightful, and designed to help the user elevate the quality of their writing. Focus on maintaining a professional and concise style suitable for business and productivity contexts. |
I have highlighted (in red for the suppression and green for the addendum) the modifications; not much, right? Yet, according to Copilot, By making these adjustments, the prompt becomes more aligned with Copilot’s capabilities, ensuring you receive the best possible assistance for your text editing needs.
After such a long introduction, you might wonder what this post will be about. Given these premises, I decided to learn more about Copilot prompts.
Learn about Copilot prompts
My first stop was at this so-called article from Microsoft: Learn about Copilot prompts. In short, to get the results you want,
prompts should include four parts:
Goal – what do you want from Copilot?
Context – Why do you need and who is involved?
Expectations – How should Copilot respond to best fulfill your request?
Source – What information or samples do you want Copilot to use?
In keeping with the prompt engineering best practices, Microsoft recommends 1) including details, 2) structuring your prompts, 3) providing positive instructions, and last, but not least, 4) iterating and regenerating.
When it comes to prompts, order matters. The order of your instructions can impact the response you get. Later parts of a prompt are likely to be emphasized more than earlier parts.Microsoft
In another article — Cooking up a great prompt: Getting the most from Copilot — I found some Helpful hints to keep in mind:
Know Copilot’s limitations: Copilot is limited to your current conversation, so give lots of details.
Be professional: Using polite language improves Copilot’s response.
Communicate clearly: Pay attention to punctuation, grammar, and capitalization.
Use quotation marks: This helps Copilot know what to write, modify, or replace for you.
Start fresh: Avoid interrupting and type “new topic” when switching tasks.
Copilot Lab
Like Anthropic has its library for Claude (see Maybe with a Little Help 2.0), Microsoft has a lot of example prompts that you can edit to make your own,
the so-called Copilot Lab. I have searched the Prompts to try, focusing on prompts to edit, and to be honest, I didn’t find anything interesting.
Copilot Lab can help you get started, with lots of examples to try or change to suit your needs.Microsoft
Some limits
There is an interesting article about limits to the size of documents that Copilot can read at once.
It appears that around 80,000 words or less is a sweet spot for Copilot to summarize a document effectively
and rewrite works best on a document that is less than about 3,000 words.
Good to know!
Keeping files you provide to Copilot within these size limits helps Copilot read and respond to your documents accurately and effectively. If your document is longer than these limits, Copilot might still be able to help, but it’ll focus only on the first part of your document. It would be like reading only the first few chapters of a book and trying to guess the rest (which is what Copilot ends up doing in this case).Microsoft
Fortunately, there are some workarounds suggested in the aforementioned article:
Here are some ways that you can use longer documents with Copilot:
Break It Down: If you have a long document, consider splitting it into smaller documents and providing them to Copilot separately. This way, Copilot can handle each part effectively.
Summarize in Parts: For long reports or manuscripts, you can try summarizing them in chunks with Copilot. Today you can do that by copying/pasting chunks into separate documents and summarizing sections separately. This helps Copilot to give you more precise and relevant responses.
Remember, you have enough time to make sure Bing is prepared – so it can help to first ask if it’s familiar with Kara Swisher and her interview style. Then add more details – who is she talking to? Then fire the prompt about the 10 questions.Frank X. Shaw
A few warnings
Here are two points (from one of the aforementioned articles) to keep in mind:
Review and verify responses you get from Copilot. Copilot is built upon Large Language Models (LLMs), advanced tools designed to predict and generate text. Occasionally, Copilot responses can include incorrect content, due to the vast and diverse nature of LLMs. Evaluate Copilot’s responses and cross-reference with trusted sources when needed.
Using the same prompt multiple times can result in different responses. LLMs are built upon neural network, which introduces some randomness. Even with the same input prompt, most likely, you will get slightly different results each time.
Reminder! Avoid putting confidential information in the chat.Frank X. Shaw
Final words
In this article, I have tried to cover some specificity about Copilot prompt engineering, including the strengths and limitations of Copilot. I provide some guidelines to get great outputs from Copilot with some extra tips (and what you should avoid or be aware of). As we say in French, “y’a plus qu’à” 3.
1 ChatGPT is also open to use by the public for free, but I could not access its features without signing up or logging in (see Maybe with a Little Help). ^
2 ChatGPT has a knowledge base that’s current up to 2021. Copilot too, but it has access to real-time data through internal tools (i.e., Bing / Edge). ^
3 Colloquial ellipsis of “il n’y a plus qu’à X” (literally “now all that’s left is to X”). In other words, now that I know all that there is to know about the topic, all I have to do is put this into practice. ^