Signal, not noise

A weekly AI-news curation pipeline for one reader.

Sift fetches the AI feeds you care about, throws away what you've already seen, clusters obvious duplicates locally for free, then makes one Claude API call to merge, categorize, score, and summarize the week — and renders a clean, dark-mode-friendly digest you can read in two minutes.

It is deliberately small: one API call per week, everything else local and free.

Why it exists

The firehose of AI news is mostly noise: press-release vapor, chatbot drama, incremental product marketing. Sift is built for a reader who wants signal — real deployments, concrete numbers, primary sources — and who would rather spend two minutes on a curated digest than an hour doom-scrolling feeds.

You describe what you care about once, in plain language. Sift scores every story against that profile and keeps only the top handful.

How it works

The pipeline is five steps, and only one of them costs anything:

1. Fetch — pull every configured RSS/Atom feed. A dead feed is logged and skipped, never fatal.
2. Filter — drop anything already seen (tracked in a local SQLite database) or older than eight days.
3. Dedup — cluster near-identical headlines locally with string similarity, so the same story from five feeds becomes one cluster. Free.
4. Rank — the single Claude call. It merges clusters that still cover the same story, assigns a category, scores importance 1–10 against your interest profile, writes a two-sentence neutral summary, and flags any story whose central claim isn't from a primary source.
5. Render & deliver — write a self-contained HTML digest (and JSON), record the run and its cost, optionally email it to you, and update this site's archive.

What you steer

• Interest profile — a paragraph describing your signal vs. noise.
• Feed weights — trust some sources more than others.
• Minimum score — drop weak stories even when the digest isn't full.
• Muted topics — tell Sift what you never want to see.

See the usage guide to set it up, and the roadmap for where it's headed.

Cost

One Claude API call per weekly run. Everything else — fetching, dedup, rendering, history, the email, this site — is local and free. A typical week costs a few cents; sift history shows the exact running total.

For AI agents

A new digest publishes every Sunday. Sift exposes a machine-readable JSON API you are welcome to scrape — politely, since it only changes once a week:

• llms.txt — how to scrape, plus the digest JSON schema.
• digests/index.json — manifest of every week (week id, date range, story count, cost, and the HTML/JSON URL for each).
• digests/latest.json — the newest digest, in full, at a stable URL.
• digests/<YYYY-WW>.json — any specific ISO-week digest (e.g. digests/2026-26.json).

Each digest JSON lists scored, categorized stories with a two-sentence summary, the source links, and a needs_verification flag. Weeks are ISO year-week (UTC), labeled by the week the digest covers.