convez.ai
AgenteAutomaçãoAvançado
kippapollo

Apollo AI News

kippapollo/ai-news

Publicação automatizada de notícias IA com agente Claude em múltiplas cadências

Ver no GitHub
InicialPluginScore · de 0 a 100

Agent que gera automaticamente resumos de notícias de IA em cinco cadências (diária, semanal, mensal, trimestral e anual). O Claude Opus busca notícias na web, lê os artigos originais e escreve resumos próprios com citações diretas verificadas. Publica páginas HTML auto-contidas no GitHub Pages.

Instalação

$npx convez-ai kippapollo/ai-news

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 2 dias
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Monitoramento contínuo de tendências em IA
  • Curadoria e resumo automático de notícias técnicas
  • Publicação web sem dependências externas com GitHub Pages

Veredito

Agent robusto que demonstra orquestração complexa: busca web, validação de data em múltiplos formatos, extração de conteúdo, escrita e geração de SVG. Exemplo maduro de automação editorial em produção, com cron local e CI/CD via GitHub Pages. Código bem documentado, taxonomia clara e implementação de filtros shareable.

README

A tradução para o português deste README ainda está na fila.

Apollo AI News

Live site: kippapollo.github.io/ai-news

An automated AI news publication. Five cadences: daily, weekly, monthly, quarterly, annual. Every edition is generated by a Claude Opus 4.7 agent that searches the web, reads the source articles, and writes each summary in its own words. Pages are self-contained HTML with no external dependencies, served by GitHub Pages.

How it works

A local cron schedule on a personal machine fires claude --print with one of five cadence prompts. The agent:

  1. Determines the time window from the run clock. Daily covers the last 24 hours; weekly covers the prior Mon–Sun; monthly covers the previous full calendar month; quarterly and annual follow the same pattern.
  2. Searches the web for AI news in that window. Each candidate's publication date is verified against the article's metadata (<meta article:published_time>, JSON-LD datePublished, <time> tag, dateline, or URL date segment). Anything outside the window is dropped.
  3. WebFetches each source URL to read the actual article body.
  4. Writes a 800–1500 word own-words summary per story, with 4–8 short attributed direct quotes for color. No verbatim reproduction of source paragraphs.
  5. Generates a unique inline SVG thumbnail per story plus a generative hero artwork for the lead.
  6. Assembles a single self-contained HTML page following STYLE_GUIDE.md (typography, palette, masthead, sticky tag filter bar, <details>/<summary> story-card pattern, light/dark via prefers-color-scheme).
  7. Regenerates index.html with the latest edition as the homepage hero and the full archive grouped by cadence.
  8. Commits and pushes to main. GitHub Pages republishes in ~30 seconds.

Coverage and tagging

Every story carries one or two category tags from a fixed six-tag taxonomy:

TagCovers
TechModel releases, research, benchmarks, infrastructure. Six sub-areas: frontier models, open-source models, agents, frameworks, research, general tech (hardware, compilers, kernels).
ApplicationProducts, features, integrations, real-world deployments.
EconomicFunding, M&A, leadership, revenue.
PolicyRegulation, lawsuits, government action, geopolitics.
SafetyAI safety, alignment, incidents, misuse, security.
TrendCross-cutting narratives and shifts.

The masthead on every edition has a sticky filter bar that scopes the page to one tag. Filters are shareable via URL hash (e.g. #filter=tech).

The publication is Tech-heavy by editorial choice. Roughly two-thirds of stories per edition fall under the Tech tag, spread across the six sub-areas. The other tags fill out a representative mix.

Repo layout

.
├── README.md            this file
├── STYLE_GUIDE.md       the design system; the agent reads this on every run
├── logo.svg             Apollo AI News wordmark plus sun mark
├── favicon.svg          browser tab icon
├── index.html           homepage; regenerated on every run
├── daily/               daily editions: YYYY-MM-DD.html
├── weekly/              weekly editions: YYYY-Www.html
├── monthly/             monthly editions: YYYY-MM.html
├── quarterly/           quarterly editions: YYYY-QN.html
└── yearly/              annual editions: YYYY.html

STYLE_GUIDE.md is the single source of truth for brand, typography, palette, masthead, hero, story-card structure, tag filter UI, date discipline, voice and prose rules, and SVG sizing. The agent reads it at the start of every run.

Forking and running your own

The cadence prompts and the cron dispatcher live OUTSIDE this repo (in the operator's ~/ai-news-cron/ directory) so the operator can iterate on prompts without touching publication history. To run your own instance:

  1. Fork this repo. Enable GitHub Pages (Settings → Pages → Source: deploy from a branch, main / / root).

  2. Replace logo.svg, favicon.svg, and the "Apollo" wordmark references in STYLE_GUIDE.md and the cadence prompts with your own brand.

  3. Set up a local cron script that invokes Claude Code in print mode against the appropriate prompt file. A minimal dispatcher per cadence looks like:

    #!/usr/bin/env bash
    set -euo pipefail
    cd /path/to/your/fork
    git pull --rebase --autostash origin main
    claude --print --dangerously-skip-permissions \
           --model claude-opus-4-7 \
           --max-budget-usd 20 \
           --no-session-persistence \
           --add-dir "$(pwd)" \
           < /path/to/prompts/daily.txt
    
  4. Wrap the dispatcher in flock if you run multiple cadences that may fire on the same day (the 1st of January fires all five), so they serialize instead of racing on git push.

  5. Make sure gh auth status is logged in for the user that the cron runs as, so the agent can git push to the repo.

A note on the content

Each story's underlying facts come from cited primary sources (link at the bottom of every card). The summary prose itself is written by the agent in its own words, with 1–2 short attributed direct quotes per story where a source's exact phrasing carries information the summary cannot. Treat the site like a press digest. Verify against the cited source before relying on a specific claim.

License

MIT. Code, design system, and template prose are free to fork and adapt.

Ver o original no GitHub

Bom para

Relacionados