Curated sources beat web search for niche briefs

When we shipped daily research briefs, we tried two approaches: let the LLM call a web-search tool to find fresh news in the topic, or have the user pre-pick a small set of RSS feeds, awesome-lists, and READMEs and feed those to the LLM directly. The list wins, by a lot. Here's why.

Web search is good at the front page

Web search returns whatever is popular, recent, and indexed. For broad consumer topics ("AI news"), that's fine. For a specific niche — agent runtime security, indie-SaaS analytics, edge ML inference — it returns the same five high-DA blog posts everyone else is reading. You become a content arbitrage middleman for TechCrunch.

The competitors in your niche know about those posts already. So does your audience. You sound like the rest of the timeline.

Curated lists are the niche

A niche RSS list or an awesome-list represents what people deep in that field actually read. It's filtered by someone who lives in the space. The signal-to-noise ratio is dramatically higher than search.

When the LLM has access to those sources directly, it can:

• Pick papers/tools that the casual search wouldn't surface
• Anchor every signal in a specific link
• Cite primary sources, not third-hand summaries

What we did about it

Setup has a "Curated sources" section. You paste RSS feed URLs (Hacker News, arXiv categories, niche subreddit RSS, specific newsletters) and GitHub markdown URLs (awesome-lists, READMEs, paper indexes). The brief generator pulls them all on every run and lets the model pick the strongest 5-8 items.

Adding 3-5 well-chosen sources turns the brief from "generic AI news rephrased" into "five specific things experts in my niche actually care about this week." The difference is bigger than any prompt tweak.

The asterisk

This means there's setup work. You have to know your niche well enough to pick the right feeds. That's a feature: anyone willing to do it builds a moat. Anyone who isn't shouldn't be posting daily on a niche topic anyway.