climatepulse

climatepulse started as an app: a personalised daily climate, energy, and sustainability digest. The realisation was that it didn't need to be an app at all.

The app was a real production system, not a prototype. Next.js and Supabase on Vercel: a five-phase daily pipeline — ingest from fifteen-plus RSS feeds plus scraped and API sources, full-text extraction, two-stage AI enrichment against a taxonomy of 12 domains and 108 microsectors, a personalised briefing per reader, a generated morning podcast. Cron routes, row-level security, entity resolution, a vector memory of prior coverage. It worked. It also had to be kept alive.

So it retired into a skill. The same product — fetch the configured sources, dedup, score and tag each article, synthesise a dated briefing — now runs as a reusable agent skill against a local wiki, no deployment to keep alive.

The skill kept what mattered and dropped the rest. A one-time onboarding interview tailors the source list; the fetch is deterministic; scoring self-tunes from feedback over time; the output is dated Markdown briefings plus structured per-article records any agent can read back. And because a skill is a folder, not a service, it distributes as a repo — the public version lives at climatepulse-skill.

Everything the app had to host — auth, crons, storage, email — was infrastructure in service of one briefing a day. Once an agent could produce that briefing wherever it runs, the interface was overhead. The arc from app to skill is the point: knowing when the interface should disappear.