Back to floatbird
Live · web · refreshes hourly, 9–5

news

An LLM reads the news so you can stop refreshing it. Every hour through the day it pulls a dozen world, tech, and politics feeds, distills them into the few stories that matter, and writes facts-only summaries. Then it makes falsifiable predictions with deadlines and grades them in public against the live web.
Three sections, refreshed every hour.

Why it exists

News sites ask for your attention all day and never tell you whether they read the moment right. This one keeps score: a short front page written as facts rather than takes, and a model that makes concrete, checkable predictions and owns the results.

How it works

The whole page is a single Claude run, repeated every hour from 9 to 5. One agent, driven by a packaged skill, fetches the top stories from a dozen world, tech, and politics sources, rewrites each headline so it tells you what actually happened, ranks them by real-world impact, and pulls the discussion context from Reddit and Hacker News threads.

In the same run it grades the predictions it made earlier — researching what happened with live web search and citing the source — and makes a few new bold-but-likely calls, each with a dated deadline. A running track record shows how often it has been right.

It’s a small Django app, part of the floatbird monorepo. The brief itself is one claude -p run on Claude Opus; Django just stores the result and serves the page.