I spent a day building a system that replaces two hours of Reddit scrolling every morning.

I had a problem. Every morning I'd open Reddit and X and spend way too long trying to figure out what happened in AI overnight. What shipped. What broke. What people were actually talking about versus what just got hyped. It felt like work before the work started.

So I asked Claude.ai to help me build something that would do it for me.

I didn't say "build me a Next.js app." I described my actual problem. I want to know about new AI releases. I want to catch early niche discussions before they go mainstream. I want to know which Reddit posts are worth commenting on and what angle to take. I want this in my inbox at 8 AM without touching a keyboard.

Claude.ai ran deep research on every part of it. Compared the X API (now $200/month, useless for this) against third-party alternatives. Found that Anthropic doesn't have an official RSS feed but a community GitHub project generates one hourly. Checked Vercel cron timeout limits and recommended GitHub Actions instead. Drew the full architecture as a diagram. Wrote the implementation plan with phases.

I opened Claude Code and started building.

What it actually does

Every morning at 8 AM, GitHub Actions runs a pipeline. It fetches from five sources simultaneously: Reddit RSS (no API key needed), SocialData.tools for X keyword searches, Hacker News via the free Algolia API, Exa for semantic discovery of emerging tools, and RSS feeds from OpenAI, Google DeepMind, and Anthropic.

Here's what the whole system looks like:

That's roughly 150 items. Claude Haiku 4.5 scores each one on two axes: how relevant is this to my specific stack and goals, and how interesting is this to any AI-curious developer. Two scores, one API call per item, Batch API so it runs cheap.

The top 20 personal items land in my inbox with suggested comment angles. The top 10 public items go to a dashboard anyone can read.

Total pipeline cost: about 25 cents a day.

The build

Two sessions. Yesterday and today. I was not in a flow state the whole time. Reddit's API registration page has a bug where the policy checkbox doesn't render, which sent me in circles for 20 minutes. A Twitter date field was named tweet_created_at not created_at. The Batch API rejects custom IDs with colons in them. GitHub Actions timed out because scoring 300 items sequentially takes longer than 20 minutes.

Every one of those bugs got caught and fixed within the same session. Claude Code found the date field mismatch by logging the actual API response. It caught the ID collision. It diagnosed the timeout, capped the item count, bumped the limit, and retriggered the run.

The part that still surprises me

The architecture came from a conversation. Not from me knowing what SocialData.tools was, or that Exa indexes X posts semantically, or that the Anthropic RSS feed lives on a community GitHub project. I described what I wanted to read at 8 AM and Claude.ai reverse-engineered a stack that could produce it.

Then I deployed to Vercel. Four clicks, a domain setting, done. digest.fieldnotes-ai.com is live.

In a previous life this project would have taken me a week minimum. That assumes I already know how to code. If I didn't, it wouldn't have happened at all.

I'm not saying any of this is perfect. The scoring is still too generous on Claude computer use articles -- yesterday was a big Claude news day and six of my top items were the same story from different sources. That's a tuning problem, not a pipeline problem. I'll fix it.

But the pipeline ran. At 8 AM tomorrow it will run again, automatically, and I won't have to think about it.

That's the part that feels like magic right now.

It's live at digest.fieldnotes-ai.com Still rough around the edges -- the scoring needs tuning and the design needs work. But it runs.