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
significance	How big is this, with you removed from the picture?	the world
relevance	How directly does it touch you?	you

security-news extends the base profile, overriding the rubric, the decision table, and the shape of interests.toml. Axis IDs are shared vocabulary across profiles, so a tuned rubric never makes the accumulated corpus incomparable with itself.

The agent never writes a priority. It scores the axes; a script applies the decision table. A rule that cannot be violated does not need to be enforced.

Install

Download news-digest-vX.Y.Z.zip from Releases, then register it:

• In the app (Claude Desktop, claude.ai, mobile) — add the zip from the skill settings (Customize → Skills). Prefer this route; it survives changes to where skills are stored on disk.
• Claude Code — unzip news-digest-vX.Y.Z.zip -d ~/.claude/skills/, or into a project's .claude/skills/ for a project-scoped install.

From a checkout:

make install

That builds the release zip and unpacks that, so what you run is what a release ships — a packaging defect breaks your install rather than reaching users. make install DEST=/path/to/skills installs elsewhere; make uninstall removes it.

Usage

/news-digest --repo ~/path/to/your-corpus
/news-digest --repo ~/path/to/your-corpus --since 2026-08-01 --dry-run
/news-digest --repo ~/path/to/your-corpus --source jpcert

Argument	Default	Meaning
--repo <path>	current directory	The corpus repository
--since / --until	previous day 00:00 / now	Collection window
--source <id>	all sources	Restrict to specific sources (repeatable)
--dry-run	—	Neither notify nor commit
--no-post / --no-commit	—	Suppress one or the other

A ready-to-copy corpus template is in news-digest/examples/corpus/. news-digest/references/setting-up-a-corpus.md is a runbook the agent can follow with you — the interests, feeds, and destination are yours to state; the mechanics are not. news-digest/references/scheduled-operation.md covers making the run recur every morning.

Requirements

Python 3.11 or later, and nothing else. The scripts use only the standard library — no package installation, no service, no database. Feed collection needs no credentials; sources that require a token reference an environment variable by name, never a literal.

Sending a digest uses whatever messaging tool is available to the agent at run time. The configuration names a destination, not a transport, so there is no dependency on any particular local CLI or MCP server; if nothing suitable is available, the run completes and says so.

What it does not do

No web UI. No full-text archive — only summaries of what you marked as must-read are kept. No crawling beyond the feeds you registered and the articles you chose to read. Feeds and article bodies are treated as untrusted throughout: article text reaches the agent inside nonce-delimited tags, and URLs appearing inside an article are never fetched.

Development

make check      # structural validation (= make test)
make install    # copy the skill to ~/.claude/skills/news-digest
make package    # build dist/news-digest-vX.Y.Z.zip

Design records are in docs/en/adr/ (Japanese: docs/ja/adr/).

日本語版は README.ja.md にあります。

License

MIT — see LICENSE.