Reading system
Paper Route
- Type
- Reading system
- Status
- Active
- Built with
- Python, Claude Code, RSS, Exa, Readwise, Resend, Vercel
- Replaces generic feed triage with a twice-weekly digest tuned to a written taste profile.
- Combines mechanical filtering with an agentic editorial pass instead of asking automation to do every judgment task.
- Publishes to a static Vercel page with one-click Readwise save links and optional email delivery.
Why I Built It
Paper Route is an attempt to make the internet feel smaller and better. I did not want another infinite feed or a higher-throughput inbox. I wanted a reading list that understood the kinds of things I am actually glad to have read.
The core idea is simple: write down my taste clearly, gather a manageable candidate pool, and let an agent do a slow editorial pass against those preferences. The point is not to stay informed about everything. The point is to find the 10 to 15 pieces that are worth carrying into a week.
How It Works
The pipeline has four stages. First, it discovers candidates from a small RSS spine, Exa semantic searches, and a few targeted web searches. Then Python scripts dedupe against a committed seen-state, apply hard excludes, and produce a compact candidate file.
The interesting step is the judgment pass. A scheduled Claude Code session reads the runbook and my preferences.md file, fetches finalist text where it can, and writes a digest issue with short "why you care" notes. A build script turns that digest into a static site. Another endpoint can save items to Readwise, and an optional email can send the top three.
The system is intentionally boring at the infrastructure layer: static files, committed history, and Vercel deployment. The taste profile is where the product lives.
What I Like About It
Paper Route is a good example of using agents as judgment amplifiers rather than just task runners. The machine is not deciding what I value from scratch. It is operating inside a set of preferences I can inspect, edit, and improve over time.