It's January 5th, 2026. First day back after the holiday break. Time to take stock of where BuzzRAG is and figure out what comes next.
Why I Built This
I built BuzzRAG because YouTube's algorithm didn't fit how I wanted to consume content. I subscribe to channels for a reason—I want to see their newest videos in chronological order, not whatever YouTube thinks will maximize my watch time. BuzzRAG gives me that control back.
The first thing the feed does is filter out Shorts. They're still in the data, but I don't get bombarded with them. This alone makes the experience dramatically better.
The Current Workflow
Every day, I check /admin/feed for the latest videos from my subscriptions. Each video card has a few buttons—a sparkle for transcription, a bookmark (that I never use), and a summary button for channels on auto-transcribe. The "watch later" feature is completely outdated and unused.
Here's what I've learned: I prefer a selective approach. I used to have many channels on auto-transcribe, which meant parsing through endless summaries and articles. Now I use the transcription button more deliberately. Less noise, more signal.
After exhausting new videos in the feed, I click over to /admin/stories. This is my favorite part of the system. The summaries are excellent—created by capturing the full transcript from YouTube captions and sending the first 15k tokens to an LLM for summarization.
Long videos are tricky. 15k tokens covers roughly 1-1.5 hours of video. Most videos fit within this window, and longer podcasts usually front-load their best content anyway.
The Numbers
Let me be real about where we stand:
- 109 articles published
- 160 approved but unpublished
- 97 in escalation (failed 2+ review attempts)
Unpublished articles are generated about 2.5x more than published ones. Many published articles come from escalations that I manually approve or regenerate with editor notes. The main feedback? Writers not using authentic voice, boring phrases, and occasional factual issues.
This is the core tension: I want to use better models like GPT-5 or Claude 4.5, but the cost is prohibitive when 5-6 generations go unused for every one that publishes.
What's Working
The summaries. They're genuinely useful for deciding which videos might make good articles. I'm starting to think summaries might be the real value here, not the articles themselves.
The personas. I find myself hoping a certain writer gets assigned to a story. Sometimes I'm not happy with the match because of their backgrounds, and that feels... real. It's subtle but embedded in the fabric of the magazine.
The articles themselves add something to the conversation. They look at content from the writer's perspective. That was always the goal, and it's actually working.
What's Not Working
The wasted generation. An article in escalation has 3 prior attempts—that's 300+ API calls going nowhere.
The workflow coupling. Right now, creating a summary automatically triggers article generation. But I often want the summary without necessarily wanting an article. Maybe a button on the summary card should queue for article generation instead?
Tracking what I've seen. I watch how many hours ago a video was released to track what's new. It works for now, but it's not scalable for generating a daily magazine with minimal human interaction.
Features I Don't Use
The publishing calendar is complicated and I'm still trying to understand the workflow. Twitter is our only distribution channel right now. Google SEO flagged us initially, but after adding AI disclosures and resubmitting, the warning was removed. Still no indexing though.
The outreach tools bring me face-to-face with content creators, and I'm not ready for that yet. There's nothing on the BuzzRAG front-end to attract them. Maybe it's fear. I should use these tools more.
The Small Wins
I showed BuzzRAG to my friend Johnny. He said it was remarkable. I said anyone could do this. He said he doesn't see people doing things like this a lot.
That gave me hope.
We have 2 Twitter followers. One is probably a bot. But still—this thing is live. There are scheduled tweets, newsletters, and real articles being published. A few months ago this was just a video feed reader.
What's Next
A few things on my mind:
-
Decouple summaries from articles. Let me generate more summaries without triggering wasted article generations.
-
Add manual video imports. Sometimes I see a video from a year ago that I want to write about, but it'll never enter the system under the current structure.
-
Reduce generation waste. Either improve the success rate or get smarter about which videos are worth generating articles for.
-
Figure out the creator angle. How do we make BuzzRAG something creators want to use to help get their message out?
The Real Talk
I sat down to write this on the first day back from holiday, partly to capture the state of things, partly to create a record of building this in public.
BuzzRAG started as a simple idea: make my YouTube subscriptions more useful. Now it's a live magazine with AI writers, scheduled social posts, and a newsletter system. 109 published articles. It's really happening.
Let's see what 2026 brings.
This is the first post in our "Building in Public" series, where I share raw thoughts on building BuzzRAG—the decisions, the struggles, the wins.