Ari's Reader: My own Google Reader, about 100 prompts later

I just released a new feed reader, called Ari’s Reader. It took me about 100 prompts.

A feed reader is a place where you’re in charge of the algorithm and which information comes at you. There’s far more to curate in 2026 than there was when Google Reader shut down in 2013. My tech reading had shrunk to Hacker News, Ars Technica, and the SomethingAwful forums, and I was spending a lot of time on Reddit and YouTube.

Ari’s Reader puts all blogs and all of your social media aggregators in one list sorted by “Magic”: Claude Haiku scores each new item against interests you write in plain English and says why in a few words. My first interests prompt filled the top of the list with coding agents, which says more about me than about Haiku.

Ari's Reader showing All items sorted by Magic: a sidebar of folders (Aggregators, Blogs, News, Video) and a list of stories from Hacker News, Reddit, and other aggregators, each with its score, comments, "also on" chips, and a short purple reason such as "UEFI firmware code-execution security flaw".

A Discover page turns the aggregators I follow into a source of new blogs, ranked by what I open and like there.

Here’s the launch video. An agent made that too, music and all.

“Personal software” is here

I think coding agents can now build highly functional software for an audience of one (plus a few friends), along with everything around it. Agents wrote the 528-line design doc, the README, the deploy config, and a post-deploy smoke test. Another agent made the video from footage of a sandboxed copy of the app, picked a CC-BY track (Kevin MacLeod’s “Voxel Revolution”), and cut it to the beat. The orchestrating agent even wrote a build report.

What it cost:

About 100 prompts from me, many sent while I used the app for real.
About 21k lines of Go, JavaScript, and CSS, including 4.7k lines of tests.
About $206 of Claude Opus tokens at API list prices (my subscription covered it).

Guard rails

Most of my prompts were about process, not code:

Questions first. I asked for two rounds of questions, product and then engineering. They settled 12 decisions before any code.
A contract. The design doc listed every endpoint and a 13-point acceptance checklist, so a backend agent and a frontend agent built in parallel without talking to each other.
Reviews. /simplify and /code-review on each codebase found 20 bugs, all fixed with regression tests. One was an account-takeover path through unverified Google emails (!!).
An independent judge. A fresh agent that saw only the checklist graded the app: 11 pass, 2 partial, 0 fail.
Me, using it. I imported about 115 feeds mid-build and found what no reviewer could: Reddit rate limits and a keyboard-scrolling regression.

They worked better than I expected. Getting it live took 3 deploy-config fixes. Since launch I’ve made only 2 small fix commits: a 2-line change to the home screen label, and moving the job that wakes the free Render server off GitHub Actions, which ran it only 4 times on day one. One open issue: YouTube’s channel feeds have been returning 404s on and off (58 of the app’s 67 on launch night), so a YouTube Data API fallback is on my TODO list.

Two phone screens: the All items list in Magic order, and a full-screen article with Like, Dislike, Star, Open, and Comments along the bottom.

What personal software have you built for yourself? I’d love to hear about it. And if you’d like to try Ari’s Reader, send me a quick email and I’ll invite your Google account.

Thanks to Nick Snyder for inspiring the app with his list of awesome software development infra people to follow, and Scott Bailey for trying it out.