Small Wins
A good-news site in English and Portuguese, found and drafted by an AI pipeline, approved by a person.
- Next.js 16
- Supabase
- Anthropic API
- Tailwind 4
What it is
Every hour a pipeline reads the day’s news, picks out the good stories, scores them and writes short summaries in English and Brazilian Portuguese that credit and link the original. A person approves every story before it goes live. Five a day, no doom.
The pipeline runs end to end today, from fetching feeds to choosing a licensed image for each draft. The review back office and the public site come next.
It also feeds Unscene: when Unscene hides a story about someone you’d rather not see, it shows a Small Wins story in its place.
What’s interesting
- A model per step. Haiku does a quick pass over headlines in batches, Sonnet reads and scores the survivors, Opus writes the drafts. Everything runs through the Message Batches API with prompt caching.
- Two scores, not one. Warmth (how much it makes you smile) and Impact (how much the world got better), plus flags for press-release puff, unverified claims and sad stories with a hopeful ending.
- Measured, not vibes. My approvals become labels: half feed the prompt as examples, half grade the pipeline in evals.
- A hard budget. Every call’s cost is recorded. Each batch is sized to what’s left of the day’s budget, and the pipeline stops itself at about $10 a month.
- Safe to run twice. An hourly tick does a bounded slice of work and stops before its time limit. Overlapping ticks, retries and batches that never answer are all handled, so a failed run costs nothing.
- Image licensing. A source photo is used only when its page or file says the licence allows it; otherwise the draft gets a credited stock photo.