From Inbox Overload to Audio-Friendly Insights: Building the motyl.dev News Engine

TLDR:

Tired of "walls of text" and the frustration of TTS engines choking on code listings, I built a custom pipeline to process technical newsletters. By leveraging Gemini, Claude, and an MCP server, I now turn my Gmail backlog into audio-friendly Markdown articles for my site, perfect for listening during my morning coffee.

The Problem: Newsletter Fatigue and the "Wall of Text"

The "News" section on motyl.dev didn't start as a grand design; it was a desperate solution to a very specific problem. Like many in tech, I subscribe to numerous newsletters. But technical content is demanding. It’s full of code snippets, complex listings, and dense "walls of text" that require significant mental energy to process.

I am someone who consumes knowledge primarily through audiobooks. I prefer that flow. However, my inbox was becoming a graveyard of unread technical articles. Even though the content was interesting, I felt a massive mental resistance to sitting down and scrolling through it all. My frustration grew as the pile of unread mail got higher and higher.

The Tooling Dilemma: Edge vs. ElevenReader

I’ve long been a user of Microsoft Edge, which might not be the most popular browser, but it has one "killer feature": a brilliantly executed Text-to-Speech (TTS) engine. It reads articles in almost any language using incredibly natural-sounding voices—and it does it for free.

I looked into alternatives like ElevenReader by ElevenLabs. While the quality is top-tier, the pricing model felt like a disaster for my needs. When a TTS service costs as much as a Storytel subscription (which includes the actual content/books), it feels overpriced. We are living in an era of "Subscription Fatigue" where everything is a "SaaS" or an "as-a-service" model. Whenever possible, I look for ways to achieve similar quality using existing or free services rather than adding another $20/month to the pile.

The Disaster of Raw TTS on Technical Content

The real breaking point was trying to use standard TTS on raw technical newsletters. It was a catastrophe. The automated voice would try to read every single character in a code listing—hashes, brackets, obscure syntax. It was unintelligible and mentally exhausting.

I found myself procrastinating even more. I didn't want to unsubscribe because I wanted to stay up-to-date, but I couldn't bear to read or listen to them in their raw form. The "pile" kept growing along with my frustration.

Enter AI: From Manual Prompts to Automated Flow

I started experimenting with AI to bridge the gap. Initially, I created a prompt for ChatGPT, Google Gemini, and Claude. The goal was to have the AI "look" at the newsletter and rewrite it in an "audio-friendly" format—skipping the code listings but describing what they did.

Interestingly, I noticed that Gemini handled paywalls much better than Claude. While Claude would sometimes refuse to access a Substack link due to a paywall, Gemini had no such qualms.

However, manually copying and pasting text on a phone was clunky. I needed a system.

Evolution 1: The GitHub Repository

I first built a system called newsletter-ai. It used LLMs to connect to my Gmail (keeping credentials secure, of course) and generated Markdown files. Suddenly, I realized: My blog, motyl.dev, is already built on Markdown. If I pushed these generated summaries to my site, I could use the Edge "Read Aloud" feature to listen to them perfectly.

Evolution 2: The MCP Server and CLI

The API costs were roughly $1 per newsletter. That’s both cheap and expensive; over a year, it adds up. To optimize, I started using tools like Claude Code and Gemini CLI.

The final piece of the puzzle was creating a local MCP (Model Context Protocol) server. This server runs locally, connects to my Gmail, finds newsletters based on specific patterns, and feeds the content directly to the LLM. Additionaly it saves some tokens by using node article-extractor library to scrape content of articles. MCP is able to provide extracted content, number of newsletters in mailbox, list of newsletters, and more. This MCP is part of newsletter-ai project now.

The New Workflow: Morning Coffee and Curation

Now, my morning routine is simple:

1. I run a single command in my CLI.
2. The system fetches new newsletters, generates audio-friendly summaries, and deletes the original emails from my inbox.
3. I listen to the generated articles while drinking my morning coffee.

I even added a "bookmarking" feature for the most interesting articles. This is the "Aha!" moment: this filtered list of high-value, front-end, and productivity insights isn't just for me. This curated stream will eventually become the basis for my own newsletter, which you can subscribe to for the "best of the best" technical content.

The system I described isn't just about saving money; it's about reclaiming time and turning a source of frustration into a streamlined source of knowledge. It is example of leverage of AI and automation to improve quality of life.