← Back to the Metis case study
Metis · Personal AI assistant · Sample issue

AI Brief: a daily AI news digest, ranked for one reader

Every morning at 6:06, Metis pulls the past 24 hours from lab blogs, Hacker News and Hugging Face Daily Papers, removes duplicates, and writes a short ranked brief for each reader: their own language, sections and priorities. Below is a real issue from Saturday, Oct 10, 2026. Switch the reader language to see the same issue in Chinese, the language it was originally sent in, or in Korean.

13 sourcesRSS, APIs, one scraped page
1 model callper reader, per day
≈ $0.05per issue at API prices
3 languagesChinese · English · Korean
Daily since 10/09no failed runs
Reader language
μ
Metis <[email protected]>
Metis AI Brief · 10/10 · Sat 6:06 AM

How an issue is made

The expensive part, judgment, is the only step that uses a model. Everything else is plain code, so the format stays identical every day and each issue costs about five cents.

Fetch

RSS / Atom feeds, Anthropic's news page, Hacker News (100+ points, AI keywords) and HF Daily Papers. Each source is fetched once for all readers.

Python

Trim & dedupe

Per reader: keep their source groups, drop links already sent, cut every item to title, summary (≤400 chars) and link.

Python

Rank & write

One call per reader with their audience, sections, language and priority rules. Returns structured JSON, not prose.

Claude Sonnet

Render & send

A fixed HTML template turns the JSON into the email and sends it from a dedicated mailbox at 6:06 every morning.

Python

The Chinese version is the issue exactly as sent. The English and Korean versions carry the same items, rankings and links, translated for this page in the pipeline's two other output languages. Metis runs locally on a Mac Studio and uses Claude through the owner's subscription. The same pipeline serves a second reader in Korean, with an extra motion- and graphic-design section and plain-language explanations of AI jargon.