Catch off-brand writing before it's sent

All right, today we're going to be building a proactive Superhuman Go agent. These agents are different than any agent you've ever seen before. If you're accustomed to agents with ChatGPT or Claude or any other software, no other agents I've seen can do this. So follow along with me as we build a proactive agent that acts before you even ask.

To set the context, we're going to be building an agent called Brandman. Imagine that I am an employee at a B2B software company and I'm writing to customers. Well, I don't want to just write however I want to write. There's certain brand and tone guidelines that I need to follow in order to be on brand. So we want to ensure that we enforce that across all of our employees across our company.

So how we're going to do that here is we have a brand guideline. I have this document that tells my employees how to write. But the problem is there's a lot of information in here and my employees don't always follow it. Sometimes they just go into their email and send off emails. So how are we going to get people, while they're writing, to follow these guidelines?

What we're going to do is go back to Superhuman Go and create an agent called Brandman. Our description here is to help ensure that our customer support and customer-facing roles communicate on brand. The descriptions are important. They do help other agents know when to orchestrate or use this agent on your behalf.

So I'm going to give it some instructions. First, I'm just going to give it a role. The company that we're pretending to be is Fernway. So this is Fernway's brand voice editor. You watch what the user is writing and help them match the Fernway brand voice.

And this is how we give it our brand guidelines: I could just put the entire brand guidelines into the instructions here, but it's hard to change. What if your brand guidelines change over time? The beauty of keeping them inside of this document is that your entire team can come and collaboratively update it whenever they need. And we're going to ensure that our agent is dynamically referencing that live document.

So how I'm going to do that is I'm going to say the guidelines live in a Coda document. Always use the Coda or Superhuman Docs MCP to read them before giving any feedback. In order to make this agent faster, I'm going to specifically tell it where to find that information. So I said, please use the document ID and use the page ID. If you're unsure how to get those, you can actually get those up here: doc ID, or go to the page itself and you should see a copy page ID. If you don't see those, go to the comments down below. There's some more information about how you can find that for your account. It may not have that turned on, so check out the description below for more resources.

So I've got this here now, and back in my instructions I'm going to tell it to use the content read tool, request markdown content. I'm going to tell it when to act specifically. I'm going to tell it how to give feedback. This is important because our agent will proactively, as you're typing, come in and tell you what you should do differently. We don't want the agent to just be like, you suck, write differently. We want it to give specific feedback, and not just feedback, but we're going to suggest a rewrite of that phrase or sentence. So it's not only going to tell you which rules you're breaking, it's going to be able to instantly change your writing into an on-brand style.

All right, so here we go. We have triggers down here. It's a proactive assistance while writing. Let's delete that though and show you how we actually built that. So we're going to go to add trigger and there are a lot of triggers here. They are growing at the time of recording. These are the ones available, and we're going to choose the proactive assistance while writing. It is active. You can change the color here if you'd like. Purple. Red is like warning, warning, warning. I'm going to choose purple because I like it.

Where would you like this agent to help you? I'm going to say email, docs, not in coding because I'm not talking to customers while I'm coding. I'm not doing it in social media and we're not doing it in my search, my browser. So I don't want it to trigger just all the time. I want it to trigger most likely when I'm talking to a customer. If I find that this triggers too often, what you can do with your proactive triggers is come down to the custom instructions and give it more information to restrict when it's going to trigger, or to make it trigger more often. But I think this is fine for now.

So what we're going to do lastly is turn on the Superhuman Docs MCP. I've asked it to use the MCP in order to grab that information. It won't be allowed to, or able to, unless you turn this on. So let's turn that on. And this says always ask, for writing. We're not having it write at all, so it doesn't need to do that, but we can keep it always ask for now.

All right, and with that, we now have our proactive agent ready to go. I'm going to go ahead and save it, and then let's move on.

Now the agent is saved and I'm going to move into email here. I'm in Superhuman Mail right now, but this would theoretically work on any email client, whether you're Gmail or Outlook or wherever else you are, it should be able to work. So I'm going to go ahead and first show where Go is. This is my desktop version of Go. You'll see I can open Go here and just chat with it. Or the other part down here are the agents for writing. You should be able to see them while it's working.

So I've got a message to a customer. Let's say, message to customer: I'm going to say, "Sorry for the inconvenience. The login issue impacted some users this morning." These first things that you're seeing are Grammarly suggestions, but let's go here and see what's happening.

You'll see Brandman is currently checking, and it says it has an issue. If you see this, there might be an issue with you here: page ID appears invalid. So it's giving us some information about what might be wrong in our configuration.

All right, so it's giving us the information that the page ID is invalid. So we can actually go back, and it should be able to work without that. So let's go and take the page ID out, and I'll say, "There is only one page in this document, read that page and use its guidelines to provide all feedback." Let's go ahead and save that.

All right, Brandman has been saved. Now let's go back here and we're going to see Brandman checking out our writing again.

And there we go. You see five issues. This will not normally show up unless I have it open, but you'll see there is our purple underline. It says, "Sorry for the convenient inconvenience." If I hover over it, you'll see that it's coming from Brandman, and it's giving me a rewrite: "This morning, some customers had trouble logging in."

Not only does it give me the rewrite that I should do, but it also gives me some information: this word breaks word choices. The guidelines say not to write "sorry for the inconvenience" and to say what happened and what we're doing about it instead. So this is where I think these agents are great, because again, it's not just asking me to go rewrite it or copy and paste from some other platform. I can simply click and it's done for me. The login issue, instead of "resources" I should say "team," and instead of "users" I should say "customers."

So let's try one other and see if it will trigger again for us. Let's go. I'm going to grab from my clipboard. I'm going to paste in an email. This is for a launch of product email to all my customers. So I've got my information here, and you'll see right there. It took about maybe three seconds that time, but it's starting to give me a ton of different information: inclusive greeting, "everyone." I want to say change "9am PT."

So all of this information is coming from my proactive agent. What is good about this is oftentimes your employees may not have the due diligence to go check brand guidelines, come back, synthesize their writing with that, and understand what's out of sync. I also don't have to take this and copy and paste it into ChatGPT and ask it to analyze against brand guidelines and paste in the guidelines. This agent will simply work for my whole company proactively to give me the writing that I need to do.

So those are proactive agents. There are so many types that you can use, because you can use other connectors at the same time. If you have any questions about proactive agents, or if you just want more resources and information on them, head to the description below to see them.