AI news

Collect the new AI stories, explain them in plain PT-BR, and (on request) turn the best ones into short-video scripts. The fetcher does the scraping; the model only reads, ranks and writes.

Sources

Both robots.txt allow crawling (Hermes asks crawl-delay 1, the script respects it). Neither has RSS. Stories are publisher claims: benchmarks and prices are "segundo X", never stated as fact.

Output location

~/.oficina/news/ (override with OFICINA_NEWS_DIR). Never inside a repo: this skill ships in a public repo and the briefings are personal.

• seen.json – URLs already delivered (the next run brings only new ones)
• YYYY-MM-DD.md – the briefing
• YYYY-MM-DD-shorts.md – scripts, when asked

Step 1: fetch

python <skill-dir>/scripts/fetch_news.py --days 2 --pages 2 > <scratch>/news.json

Flags: --days N window, --pages N Hermes pages (20 items each), --all ignore seen.json (re-run the same day). Read only count, errors and the items; do not dump the JSON into the reply. count: 0 means nothing new: say so and stop.

Step 2: read before writing

For each story kept (max 10), open the article page only when the summary is empty or the story is a ranking candidate. Prefer the original publisher link when the page has one. Do not invent numbers absent from the page.

Rank by what matters to a dev who uses AI daily: new model or price change > tool/API a dev can use today > research with practical effect > funding/corporate.

Step 3: briefing (YYYY-MM-DD.md)

# IA hoje – DD/MM/AAAA

## Destaques
### <título em PT-BR, sem clickbait>
**O que é:** 1-2 frases.
**Por que importa:** 1 frase concreta (custo, o que destrava, quem afeta).
**Fonte:** [Publicador](url) · data

## Rápidas
- <uma linha por notícia restante> – [fonte](url)

3-5 destaques, the rest in Rápidas. Plain PT-BR, technical terms kept in English when that is how devs say them (benchmark, open-weights, token). After writing, reply with the path and the destaque titles only.

Step 4: shorts (only when asked)

Pick the 3 stories with the clearest "so what" for a general tech audience. If the yt-script / yt-shorts skills are installed, follow their hook and retention rules; otherwise use this format per story:

## Short N – <título>
**Gancho (0-3s):** frase que para o scroll, com o fato mais surpreendente.
**Contexto (3-15s):** o que aconteceu, em linguagem de gente.
**Virada (15-40s):** o número ou consequência que muda algo para quem assiste.
**Fecho (40-55s):** opinião curta ou pergunta para comentário.
**Na tela:** 3-4 textos de apoio / b-roll sugerido.
**Fonte (descrição):** url

Rules: 110-150 words of narration per short (≈45-55s), one idea per short, no "neste vídeo", no invented quotes, say "segundo <fonte>" for claims. Do not reuse the source's wording or images: write original narration and credit the link in the description.

Adding a source

Add a generator to scripts/fetch_news.py yielding {url, title, date, source, summary} and register it in the loop in main(). Check the site's robots.txt first and prefer a sitemap or feed over HTML.