securesein.github.io：一个部分由 AI 自动撰写发布的个人 AI 博客

securesein.com — a personal AI blog, and the pipeline that partly writes it.

This repo is two things in one: an Astro static site, and a set of Python scripts that read AI news/research feeds, score what's worth covering, and — for some sections — draft and publish posts with no human review at all. What ships without review, what's directed by a human with a model drafting, and what's fully hand-written is disclosed on every post; the full story is on the site's About and Colophon pages.

Layout

src/                  Astro site: pages, components, the content-collection schema
src/content/blog/     Every post, as markdown with frontmatter
config/                Budgets, the reading-interest profile, and the pipeline's other tunables
scripts/
  core/               Feed fetching, dedup, scoring, the ledger, LLM calls, draft writing
  channels/           One module per publishing lane (releases, research, security, benchmarks)
  telegram/           The review/digest/feedback flow over Telegram
  papers/             Reader — a separate, read-only arXiv digest (see scripts/papers/__init__.py)
.github/workflows/    CI: one workflow per channel, plus digest/feedback/deploy

src/content.config.ts is the single source of truth for what a post is allowed to look like (section, format, up to three topics, an authorship record); scripts/core/frontmatter.py is what writes to that shape from the Python side. The two have to stay in sync by hand — there's no generated client between them.

Running it yourself

This is a personal, opinionated setup, not a generic templated product — forking it means adapting the config to your own taste, not flipping a switch. A few things to know going in:

• The Python pipeline and the Astro schema are tightly coupled. The taxonomy (which sections, formats and topics exist) lives in taxonomy.json and is read by both the site and the scripts. Change one without the other and posts stop validating.
• Scheduling is external. GitHub Actions' own schedule trigger isn't reliable enough for the timing some of these channels need, so every scheduled channel is workflow_dispatch-only and triggered by an outside cron service hitting https://api.github.com/repos/<you>/<repo>/actions/workflows/<name>.yml/dispatches with a GitHub token. This repo was built against cron-job.org; any scheduler that can make an authenticated POST works.
• Secrets. Set these as repository secrets:
  • OPENAI_API_KEY — drafting and scoring
  • TELEGRAM_BOT_TOKEN, TELEGRAM_CHAT_ID — the review digest and feedback loop
  • TELEGRAM_PAPERS_CHAT_ID — optional, only for the separate Reader digest (scripts/papers/)
• Tune before you turn it loose. config/interest_profile.yaml (what's worth writing about), config/budgets.json (how much it's allowed to publish and spend), and taxonomy.json (the vocabulary) are the levers. The defaults reflect one person's interests, not a neutral baseline.
• Local dev: npm install && npm run dev for the site; pip install -r scripts/requirements.txt for the pipeline. npm run build runs the content-collection schema as a validation gate — a malformed post fails the build, on purpose.

None of the content under src/content/blog/ is covered by any implied invitation to reuse it as your own writing — see About for what's actually true about how each piece was made.

License

MIT — see LICENSE. The code is yours to fork and adapt; the posts are still just posts.