Why a news reader is a cache problem, not a feature problem

A news reader looks like a simple list but the economics only work if you stop generating AI summaries per user. One article should be summarized once, globally, and served to all users who open it. Miss this and your LLM bill tracks your DAU multiplied by articles per session.

I have not shipped a news reader, but the caching pattern is the same one I used for the meal plan endpoint on a recipe experiment, and it keeps cost per user under a penny per day in practice.

What you actually need to build

• Feed management: Add/remove RSS feeds. Seed with 10 popular tech or news feeds.
• Unified timeline: Reverse chronological articles from all subscribed feeds.
• Read state: Mark as read on tap. Sync across devices.
• AI summary per article: One 3-bullet summary per article, generated once, cached by article_hash, served to all users.
• Starred/saved: A simple second screen for saved articles.
• Offline read: Cache the last 50 articles per feed locally so the app works underground.

No custom reader mode, no filtering by topic with AI, no notifications per keyword in v1.

The stack I use

• React Native + Expo.
• rss-parser on the backend, not the client. You do not want CORS pain on mobile RSS.
• Supabase — feeds, articles, user read state. The article cache is a global table keyed on url_hash.
• NestJS — hourly RSS fetcher that ingests new articles. Summary endpoint that returns cached summary or generates one.
• Claude Code + 11 AI agents — scaffold the feed list and article screens.

Real build time

With the boilerplate, 3 weekends.

• Backend RSS fetcher + articles schema: ~6 hours.
• Feed management + unified timeline: ~6 hours.
• Read state sync + starred: ~4 hours.
• AI summary endpoint + global cache: ~4 hours.
• Offline cache of articles: ~4 hours.
• Store submission + attribution rules: ~4 hours.

About 28 hours.

Where people get stuck

• Per-user AI summary cost: If every user triggers a new summary generation for each article, you spend ~$0.001 per article per user. A cache keyed on article_hash takes it to roughly one generation per 1000 views. Difference is orders of magnitude.
• RSS feeds that change their format: Every major feed breaks parsing at least once a year. Log parse failures and ship a fallback. Do not crash the whole timeline for one broken feed.
• Copyright on summaries: A 3-bullet summary is generally fine under fair use, but reproducing the full article is not. Always link back to the source with prominent attribution, and do not cache the article body beyond what you need to summarize.

Skip the setup

RSS fetcher, articles schema, globally cached summary endpoint, offline sync — pre-wired. The 11 AI agents scaffold the timeline and reader screens.

See pricing