\ud83d\udc68\u200d\ud83c\udf93Speak to it like it\u2019s an intern: <\/strong>Rather than using concise instructions as you might with a search engine, or assuming a shared understanding as you might with a colleague, imagine that Copilot is smart but benefits from extra context and really explicit instructions to break down tasks to understand what you want it to do and how. <\/p>\n\n\n\n \ud83c\udfad Your tone may influence the AI \u2018persona\u2019:<\/strong> You don\u2019t need to say \u2018please\u2019 and \u2018thank you\u2019, but be aware that AI generates its answers using associations between words. If speak to it like a search engine (i.e. just writing a few key terms), it will act like a search engine. If you want less superficial answers, use technical language in your prompts. If you are doing a role-play, don’t ‘break character’, or Copilot will as well! <\/p>\n\n\n\n \ud83d\udcac Don\u2019t worry too much about trying to write a \u2018good first prompt\u2019: <\/strong>In general, just tell it what you want to do, and don\u2019t give up or start again if it doesn\u2019t behave as expected \u2013 give it follow-up instructions! <\/p>\n\n\n\n \u2696\ufe0f Don’t overload a single chat: <\/strong>AI chatbots have limited ‘context windows’, which is similar to their working memory. While the amount of information they can process is growing with new updates, you still get better quality outputs if you keep chats focused on a single task. If you want Copilot to do several things related to the same topic, split these tasks up across several chats: after each chat, ask Copilot to make a summary of your conversation so far, start a new chat and paste it in, and then give it a new task.<\/p>\n\n\n\n Prefer to see these tips in an infographic style format? Go to: Specify a role and goal:<\/strong> <\/p>\n\n\n\n Act as \u2026 Your goal is \u2026 <\/p>\n\n\n\n Give it constraints: <\/strong><\/p>\n\n\n\n Use \u2026 Don\u2019t \u2026 It\u2019s essential that \u2026 <\/p>\n\n\n\n Give step-by-step instructions:<\/strong> <\/p>\n\n\n\n First \u2026 Wait for a response \u2026 Then \u2026 Wait for a response \u2026 Next \u2026 Wait for a response<\/p>\n\n\n\n If … then …<\/p>\n\n\n\n Think carefully and plan step-by-step.<\/p>\n\n\n\n Give examples and specify output format:<\/strong> <\/p>\n\n\n\n Answer in the format of the following example: …<\/p>\n\n\n\n Create a table with the following columns …<\/p>\n\n\n\n Use bullet points in your answer.<\/p>\n\n\n\n Follow up with new instructions:<\/strong> <\/strong><\/p>\n\n\n\n That answer wasn’t good because \u2026 Try again, but this time \u2026 <\/p>\n\n\n\n It’s important to be aware that you can’t completely prevent AI chat tools from ‘hallucinating’ (creating false information). This is due to the nature of how they work \u2013 they predict the next word in a sequence based on probability. However, there are things you can do to reduce the risk of hallucinations:<\/p>\n\n\n\n \ud83e\udd14 Tell Copilot that it’s okay if it doesn’t know the correct answer<\/strong><\/p>\n\n\n\n Often AI chatbots are trained to always try and provide an answer, which can lead to them making things up. Try adding something like this to your prompt:<\/p>\n\n\n\n It’s okay if you’re not certain! If you don’t have enough information or several answers are possible, explain why.<\/p>\n\n\n\n \u26cf\ufe0f If you’re asking Copilot to work with text or data, ask Copilot to extract evidence first before performing its analysis<\/strong><\/p>\n\n\n\n Asking Copilot to pull out direct quotes first, and then analyse them, helps to ‘ground’ its analysis in the real data:<\/p>\n\n\n\n First, go through the text and identify quotes about pineapple on pizza. Put these quotes in a table. Then, analyse the quotes and suggest if the author is for or against pineapple on pizza.<\/p>\n\n\n\n \ud83d\udeb6\u200d\u2642\ufe0fAsk Copilot to explain its reasoning step-by-step<\/strong><\/p>\n\n\n\n If you ask Copilot to break down its reasoning, it can help guide it towards the correct answer:<\/p>\n\n\n\n How many ‘r’s are in the word ‘strawberry’? Explain your reasoning step-by-step.<\/p>\n\n\n\n If you’re creating a learning activity with Copilot<\/a>, you may notice that it suggests the correct answers as clickable buttons above the box where you type your responses. This can be frustrating! You can’t turn this feature off, but you can explicitly tell Copilot what to put in its suggested answers. For example, if you make a Copilot Agent<\/a> that quizzes the user with multiple-choice questions<\/a>, you could include something like this in your instructions:<\/p>\n\n\n\n If you suggest responses for the user, always suggest: A, B, C. This ensures you aren\u2019t inadvertently revealing the answers to the quiz.<\/p>\n\n\n\n Copilot seems to only be able to suggest three answers at a time, so it seems to confuse it if you tell it to suggest A, B, C as the answers but to give the user four answer options. For this reason, it’s best to restrict multiple-choice quizzes to only having three options as well.<\/p>\n\n\n\n Most of the big AI companies have their own guides to writing good instructions for chatbots (this is sometimes called ‘prompt engineering’), although these can be a little technical and jargon-heavy. See, for example:<\/p>\n\n\n\n \ud83d\udd17 Prompt engineering overview – Anthropic<\/a><\/p>\n\n\n\n \ud83d\udd17 Learn about Copilot prompts – Microsoft Support<\/a><\/p>\n\n\n\nHow should I write my instructions?<\/h2>\n\n\n\n
\ud83d\udd17 Crafting Effective Prompts: A Guide for Educators (coventry.domains)<\/a><\/p>\n\n\n\nHow can I reduce ‘hallucinations’?<\/h2>\n\n\n\n
How can I stop Copilot ‘giving away the answers’?<\/h2>\n\n\n\n
Further reading<\/h2>\n\n\n\n