This page was translated from Dutch with AI.
Your first agent only sees your inbox
I am very often on the road giving training, which means I often cannot read my emails during working hours. In the introduction round of a Copilot workshop it turned out I am not alone in this, because almost everyone started talking about an overflowing inbox. One participant combined multiple roles, worked long weeks and was mainly looking for an overview of what was really important.
So I built an agent live that does only one thing: read your emails from the previous day every morning and give you a short list of what requires action. After that the participants built their own version. I actually think that is an ideal first agent, because it is small and because you already use it the next day.
I write the instruction in a normal chat first
Copilot has a wizard to build agents, and I skip that. Such a wizard wants to move on to building the agent as quickly as possible, which sometimes reduces the quality of your instruction. I also see that with colleagues who work a lot with Copilot, the power users so to speak: they always write their instructions in a chat and paste them into the agent afterwards.
What I do in the chat is tell it what I want and ask the AI to ask me questions first. I wrote that I wanted an instruction for an agent that checks my emails every day and makes a short report of what needs my immediate attention. I added some context, namely that I am often on the road, and I added that it was not allowed to execute the task itself yet. I deliberately left a few things vague, like what 'important' or 'action' means to me, because those are exactly the questions you want it to ask. It always says it is a good idea by the way, I have never seen it otherwise.
Then I answer its questions honestly, which emails count and which do not, and how I want to see the result. I chose a normal table with sender, subject and the action that is needed. At the end I ask to give the instruction for the agent as Markdown code. I copy that with the code button, so the structure is preserved, and then I paste it into the new agent.
I ask for Markdown because that is actually the notation in which language models have seen the most text. A hash for a title, a dash for a list and vertical bars for a table. You do not need to know that notation, you just need to know it exists and ask for it.
Keep reading for free
Register free with your e-mail address and read on. Your account opens the free articles in the learning space; the rest is for subscribers, from Learn Plus up.