News Digest

news-digest

A Claude Code Skill that collects your own feeds, decides what is worth reading, and compiles a digest of only that.

Feeds forwarded into a chat channel produce 130–180 items a day, which means nobody reads any of them. This skill draws three distinctions instead:

• Is it worth reading? — the significance of the event and your own relationship to it are scored separately
• Is it new? — each article is checked against the continuing stories you have already been following
• Has it gone stale? — topics that turned up again with no new facts are named as such

The third one is the point: it leaves behind a record of *why something did not need reading*.

Code here, data elsewhere

This repository holds the engine and nothing else. Your feed list, your relevance profile, and the article corpus live in a separate **corpus repository** that you own and keep private — it names what you watch and what you care about, which is not something to publish.

news-digest (public)                  your-corpus (private)
  skill, scripts, profiles,             config/  data/  digests/
  examples, tests                       .newsrc.toml

The corpus repository contains no code, and one installed skill can serve several corpora. `.newsrc.toml` is the contract between them: it declares the config and schema versions, the evaluation profile, and where digests are sent. The skill refuses to run against a version it does not support, so an engine update cannot quietly corrupt a corpus.

Profiles

The engine knows nothing about any subject area. A **profile** supplies the evaluation axes, the decision table that turns axis scores into a reading priority, the rubric the agent reads before scoring, and the digest layout.

Three axes ship as the base, kept deliberately domain-neutral because their frames of reference are orthogonal:

Axis	The question	Frame
novelty	What does this add to what is already known?	time