Takeaways

  1. My favorite part of LLM-assisted coding is educating me on tools I’ve never seen - Google Workspace user of XX years, and this is my first time seeing https://script.new and associated tooling! I’m excited to work with some small business-owner friends to use these kind of lightweight extensions to solve some of their email struggles.

  2. ChatGPT o3 continues to struggle with maintaining a set of requirements built over time. I ask it to build X, then modify it in Y way, then add Z, and in the process of adding Z we lose Y.

The project

Today’s project involves building our first Gmail extension!

🪄 Vanish Gmail is an LLM powered email assistant. It looks back over unread emails, summarizes, prioritizes, and suggests action items.

The tech stack

We’re using Google Apps script to integrate with Google Workspace, OpenAI for the LLM backend, and ChatGPT as our coding assistant.

The result

  1. I’m super happy to have built my first gmail extension! I’ve already got a ton of ideas for use cases for myself and friends here.
  2. I want to find ways of evaluating LLM performance here - prioritizing + pulling key bits from email (eventually even taking action on them perhaps? :)) is a tricky problem and even between two invocations on the same set of emails I get different (not great!) results from the model.

The code

Code here