Sample output
The actual delivered email from a real test run — not a mockup:
Full email, and the underlying archived markdown, in the repo.
The problem
Staying current on AI news means checking a dozen scattered sources every day — model releases, funding, research, policy — and remembering what you already read yesterday so you’re not served the same story twice. A naive “summarize AI news” prompt tends toward two failure modes: padding a quiet day with filler, or silently dropping a source it couldn’t reach and never saying so.
What I built
A scheduled Claude Code skill, built on Anthropic’s ai-news-digest skill and configured into a fully personalized research agent: a fixed source list with corrections for specific sources that had been silently missed before, a Google Drive archive that gives every run memory of what the last three digests already covered, and delivery as a formatted HTML email rather than a chat reply that disappears.
How it works
- Cross-run memory, not a fresh start every time. Each run reads the three most recent archived digests before searching, so it excludes what’s already covered unless there’s a genuinely new development.
- Verification tiers, disclosed honestly. Every source is marked as directly fetched, covered by search only, or blocked — never silently skipped.
- Config kept out of the skill itself. Recipients and archive ownership live in a config file on Drive, read fresh each run, so the skill file has nothing personal baked into it.
- Never ends silently. A quiet news day still produces a digest saying so, rather than no output at all.
Stack
Claude Code skill, Google Drive (archive + config), Gmail (HTML delivery via Zapier, with a native-draft fallback).