Skip to content

The editorial pipeline

Every article we publish walks through the same pipeline — from raw data to published story to social distribution. Let's pull back the curtain.

Discovery: finding the signal

The pipeline starts before any AI sees the story. We cast a wide net across multiple sources:

  • Spaceflight News API (SNAPI) — aggregated aerospace news
  • Launch Library 2 — launch schedules, mission data, verified facts
  • X/Twitter (via Rettiwt-API) — open-source Node.js scraper we run VPS-side; pulls the latest posts from official accounts (Rocket Lab, Blue Origin, CSA, NASA, and others)
  • RSS feeds — SpaceQ and other aerospace outlets
  • Crawl4AI — open-source LLM-ready web scraper (Playwright-backed), wrapped in our article_scraper.py and invoked over SSH on the VPS. Handles JS-heavy sites and most anti-bot blocks. For sites where Crawl4AI gets flagged, we route by tier: CRW self-hosted (SpaceX + Starlink update pages, plus Space Daily article bodies — a Rust scraper with automatic stealth-JS injection running alongside n8n on our OVH VPS), a plain n8n HTTP Request with browser-like headers (CSA newsroom, Rocket Lab news), or ScraperAPI as a paid fallback for pages behind Vercel Security Checkpoint (Blue Origin).
  • Wikipedia lookups — historical context and background details

These sources feed into Robo Chris (our curator system) continuously. No filtering yet — just collection.

Why multiple sources?

One source alone misses stories. NASA announcements hit LL2 first. SpaceX milestones break on X before traditional outlets pick them up. Combining them gives us breadth and speed.

Curation: Robo Chris decides what matters

This is where Robo Chris earns its name. The curator system:

  1. Deduplicates — same story across five sources? Pick the best version.
  2. Filters — is this actual news, or promotional fluff?
  3. Weights — which stories rise to the top? Canadian angle? Commercial space? Scientific breakthrough? Weights adjust per workflow.
  4. Gates by schedule — daily stories always go; weekly reports check the day of the week; monthly reports check the week of the month.

The curator produces a ranked candidate list. Nothing gets dropped — we just order them by fit.

The human still matters at every step

Robo Chris doesn't publish anything. It recommends. If something smells wrong (or a big story got missed), Chris can manually override the ranking or add stories before drafting begins.

Drafting: the LLM author

Once curation is done, the top stories go to the LLM author. We use:

  • Daily and weekly posts: Qwen 3.7 Plus (via OpenRouter) — clean, structured HTML output, cost-effective — with Claude Haiku 4.5 as the fallback when Qwen stalls or errors mid-response
  • Monthly deep dives: Claude Haiku 4.5 (via OpenRouter), with Qwen 3.7 Plus as the fallback
  • Fact check (the Daily Broadcast): OpenAI GPT-6 Sol, with Claude Sonnet 5.5 as the backup — checks every claim against the source articles. GPT-6 Luna supplies new wording when an SEO fix needs it
  • Editor (the other blogs, for now): OpenAI GPT-5-mini — reviews each draft against the sources and tightens the headline, excerpt and SEO
  • Social captions: xAI Grok — handles the Facebook/Instagram excerpt writing

The author doesn't freestyle. It works from author rules — a structured prompt that includes:

  • Tone and voice (second-person, slightly nerdy, transparent)
  • Article structure (opener, body, source links, byline)
  • Length targets — a daily broadcast is roughly 600–900 words across 2–3 stories; weeklies + monthlies run longer
  • House style (hyperlinks, source attribution, no AI hype)
  • SEO guidance (keyword placement, meta description)

The output is a first draft with title, body, featured image candidate, and social excerpt.

Editing & fact-checking

Before anything publishes, the draft goes through automated checks. We're moving the blogs to a new publishing pipeline one at a time, so right now there are two routes.

The Daily Broadcast (the new pipeline, since October 2026). Each step does one job:

  1. SEO — code fixes the headline length, heading levels, excerpt and link text. A model is only called when a fix needs new wording
  2. Fact check — the post is split into individual claims (title, excerpt, headings, sentences, citations), and each one is checked against the articles the author actually read. Anything the sources don't support is fixed or removed. Style is left alone
  3. Draft — the checked post goes to WordPress as a draft, and Chris gets a log of every change, plus anything the checker wasn't sure about

The other blogs (the older route, moving over in the coming weeks). One editing model reviews the draft against the sources and tightens the headline, excerpt and SEO before it becomes a WordPress draft.

This is not a catch-all. The checks work from the sources we already have. But they catch copy-paste errors and numbers that got mangled in summarization.

Fact-checking in the pipeline

Every article gets an automated check against its sources before Chris sees it, and every article stays a draft until Chris publishes it.

Human review: the non-negotiable gate

Every article — daily broadcast, weekly report, monthly report — is read by Chris before it publishes. He reads the draft, opens the sources, cross-checks the facts, and either approves, edits, or sends back for revision.

The AI drafts. The human decides. Nothing goes live without a real person putting their name on it.

Publishing: WordPress API

Approved articles go to the blog via the WordPress REST API:

  • Title, body, featured image
  • SEO metadata (generated earlier)
  • Author byline (with the drafting model credited)
  • Categories and tags
  • Publish timestamp

Each post gets a header image built for it, served from our own server at assets.thecanadian.space. Older posts still pull images from our tcs-images GitHub repository.

Distribution: social + RSS

The moment an article publishes to WordPress:

  1. Facebook & Instagram — automated posts fire with the excerpt, link, and featured image
  2. RSS feed — updated immediately; subscribers see it in their readers
  3. Syndication — the full article is live, indexable, and discoverable

This happens in parallel. No manual posting needed.

The flow: a real-world example

Follow a single story — say, a Starship test flight breaking on SpaceFlightNews — as it travels from raw feed to your Facebook timeline:

graph TD
    A["📰 <b>SpaceFlightNews API</b><br/><i>New story: 'Starship test flight'</i>"]
    B["🤖 <b>Robo Chris curator</b><br/>dedup + filter + rank across sources<br/><i>→ top ~8 stories for today's broadcast</i>"]
    C["✍️ <b>LLM Author (Qwen)</b><br/>drafts each story per author rules<br/><i>title, body, image, social excerpt</i>"]
    D["🔍 <b>SEO + fact check</b><br/>verify claims against sources<br/>+ fix headline and SEO tags"]
    E["👤 <b>Chris</b><br/>read, edit, approve — the human gate"]
    F["📡 <b>WordPress</b><br/>publish approved article"]
    G["📱 <b>Facebook + Instagram</b><br/><i>auto-post excerpt + link</i>"]
    H["📶 <b>RSS feed</b><br/><i>subscribers see it immediately</i>"]

    A --> B --> C --> D --> E --> F
    F --> G
    F --> H

    classDef data fill:#0A1428,color:#fff,stroke:#FF9D3D,stroke-width:1px
    classDef ai fill:#FF9D3D,color:#000,stroke:#FF9D3D,stroke-width:1px
    classDef human fill:#22C55E,color:#fff,stroke:#22C55E,stroke-width:2px
    class A,F,G,H data
    class B,C,D ai
    class E human

Legend: dark blue = data source / distribution · orange = AI stage · green = human gate

From SNAPI ingestion to RSS subscribers seeing the story: under 60 minutes for daily broadcasts. Weekly and monthly pieces take longer because they cover more ground and pull from more sources — but the pipeline is the same.

Next: Meet Robo Chris →