Email: Delete It All (part three)

Automated AI Drafts

In part one, I covered the incentives for trying this radical experiment. The objective was simple.

I enabled a custom Gmail service that monitors deleted messages every five minutes. It uses generative AI to classify each message within the ontology based on the message subject and body. When a message is detected that meets a similarity threshold, I move it into the inbox.

The Grand Email Experiment

Part two was all about a prototype that uses Google Apps Script to create a simple approach to identifying messages that were “on point”, while ignoring all others.

The OnPoint prototype is taking shape nicely. It's a Google Apps Script project. It has a simple directive - find and make visible in the inbox, any messages I care about; reject all others.

OnPoint Trainer - A Google Apps Script Function

Part three is where I successfully create an AI process that triggers whenever the OnPoint system detects a message that contains topics or tasks I need to address. If it finds anything I should probably respond to, it creates a formatted draft that accelerates my response time.

I’ve yet to see benefits and accelerated productivity from automated responses for email conversations that involve actual work topics. So, in this approach, my interest has two simple requirements.

1. Create drafts when — and only when — there’s something worthy of a response.

2. Format the draft in a way that helps me accelerate productivity.

The example message shown below is from a colleague with four topics she’d like me to respond to. The integrated AI, which triggered based on several attributes in this message, accurately determined there were questions I should address and formatted them in the manner I prefer.

Example Message with Automated Draft

This pre-formatted draft allows me to address the points and move on quickly. Eventually, I’ll make this smart(er) by integrating it with my miniME

¹

fine-tuned model project and my Calendly service.

To achieve this, I used PaLM 2 text completions. Here’s the prompt.

You are an email message response expert. I will provide you with an email message. You will generate a list of the key topics that need to be addressed. Each topics will be separately formatted on a line 
by itself prefixed with ">>> ". Add two blank lines after each topic. If there are no key topics that need to be addressed, you will output "No Key Topics" and no other text.

Example:

  Email message: Tell me about your Impertinent Substack and life in the desert.
  Output: >>> Impertinent Substack ...\n\n
          >>> Life in the desert ...\n\n

  Email message: ${threadContent}

  [output]

1. This is a very early sneak peek at a project I’m working on. For decades I have envisioned a system that would work on my behalf when I was skiing, or relaxing on a beach. It was a vision shared by many who knew that eventually, computing technology would make this possible and eventually, financially practical.

miniME is a project that camouflages my narcissistic side. It’s certainly ideal for exploring new ideas, and it also advances my productivity by making it possible to have a conversation with me while I sleep or while I’m ignoring everyone.

If there was ever a way to replicate yourself, generative AI seems to make this possible.

Imagine your most knowledgeable expert in your company. There are five meetings ongoing at the same time, and she can’t attend all of them. However, if she had a miniME, her expertise and presence could be felt in all five meetings.

Can we be in more than one place at the same time? Apparently, we’re close to saying “maybe”. The challenge is, of course, codifying knowledge. However, for this project, I realistically embraced the reality that a miniME doesn’t need to possess all my knowledge to be useful.

I capture a lot of information in a variety of systems. But for the miniME prototype (at least initially), I’m going to rely on only two - my Impertinent Substack and my website. I may add a few others as I make development progress, but for now, these two bodies of knowledge will represent the corpus of a fine-tuned model that allows you to ask questions and chat with me.

Look for some interesting posts about miniME later this Fall.