Weekly External Links

Weekly external-links refresh

Gather content published in the last N days (default 7) from Tiger Den, match it to docs pages, and write learnMore right-rail cards onto a review branch. A human reviews the branch; the skill never merges.

Background on the learnMore card: src/components/LearnMore.README.md.

Source: Tiger Den (single source of truth)

Tiger Den is Tiger Data's marketing content hub. It tracks every blog post, YouTube video, case study, whitepaper, podcast, and third-party / KOL publication as a "content item", which makes it the single source for this refresh: one place to read everything that was published.

Access is the read-only TIGER_DEN_TOKEN. The matching step uses ANTHROPIC_API_KEY. Both live in the gitignored .env locally (see .env.example) and as repository secrets for the scheduled run.

Pipeline

Four scripts under scripts/, run from the repo root. All are zero-dependency (Python stdlib only), so no install step is needed:

DAYS=7   # or the value passed in args
mkdir -p .learnmore-work

# 1. Fetch recent content from Tiger Den (token auth, paginated, normalized).
python3 .claude/skills/weekly-external-links/scripts/fetch_tiger_den.py \
    --days "$DAYS" --out .learnmore-work/tiger-den.json

# 2. Build the docs page catalog. Excludes the `reference` section by default
#    (cards never belong on function-reference pages, and skipping it roughly
#    halves what the matcher has to weigh). Captures existing learnMore links.
python3 .claude/skills/weekly-external-links/scripts/build_doc_catalog.py \
    --out .learnmore-work/catalog.json

# 3. Match + compose cards. Calls the Claude API (Opus 4.8, structured output,
#    streaming) to curate, match to pages, and draft the final card link sets.
python3 .claude/skills/weekly-external-links/scripts/match_content.py \
    --items .learnmore-work/tiger-den.json \
    --catalog .learnmore-work/catalog.json \
    --out .learnmore-work/mapping.json

# 4. Apply the drafted cards into page frontmatter.
python3 .claude/skills/weekly-external-links/scripts/apply_cards.py \
    --mapping .learnmore-work/mapping.json \
    --catalog .learnmore-work/catalog.json

# Optional: render a Markdown summary (used as the PR body on scheduled runs).
python3 .claude/skills/weekly-external-links/scripts/render_report.py \
    .learnmore-work/mapping.json > .learnmore-work/report.md

Tuning knobs (environment variables)

• MAX_LINKS_PER_CARD (default 4): hard cap on total links per card.
• ANTHROPIC_MODEL (default claude-opus-4-8): the matching model. Opus is the most careful at the editorial judgment; claude-sonnet-4-6 is a lighter, faster alternative.
• build_doc_catalog.py --exclude (default reference): comma-separated sections to skip. Pass --exclude '' to include everything.

The editorial rules (enforced in the matcher prompt)

These were set by the docs owner. The matcher applies them; the cap is also enforced in code as a safety net.

• Curate first. Drop pgai Vectorizer content (deprecated), Ghost / Agentic Postgres content (different product), and pure social one-liners.
• Hard cap: at most MAX_LINKS_PER_CARD (4) links per card, counting existing plus new. Never exceed it.
• At least 2 links per card. If a topic can't reach two, it's dropped.
• Replace, don't pile on. For a page that already has a card, the matcher starts from its existing links and folds in the new one(s); if that would exceed the cap, it drops the weakest or oldest existing links to make room, preferring a newer/more-direct link over an older one. The card's links are the final set, and apply_cards.py replaces the page's relatedPosts list with it (existing links it keeps are reproduced verbatim, so diffs stay small).
• At most 2 customer stories per card.
• No external-contributor names in labels (drop "(Justin Mitchel)"). Customer names inside a customer-story label are fine.
• No em dashes (house style). Video labels are prefixed Video: .

The card lands in relatedPosts (external-link icon, opens in a new tab):

learnMore:
  relatedPostsHeading: Related resources   # or "Watch" / "Related reading"
  relatedPosts:
    - label: "A blog post title"
      href: https://www.tigerdata.com/blog/...
    - label: "Video: a descriptive title"
      href: https://www.youtube.com/watch?v=...

Run types

Local / interactive

Run the pipeline, then review the working tree directly. Verify before handing off:

pnpm build:local        # must succeed
pnpm lint:prose         # 0 errors in your additions
grep -rn '—' src/content/docs --include='*.mdx' | grep -E 'label:|relatedPostsHeading:'   # must be empty

Leave the changes uncommitted for the human to review. Do not commit, push, or open a PR. Clean up .learnmore-work/ when done (it's scratch; never commit it).

Scheduled (GitHub Action)

.github/workflows/weekly-external-links.yml runs the pipeline weekly (and on manual dispatch), then opens a pull request against main with the drafted cards and a summary body. It never pushes to or merges main. Required repository secrets: TIGER_DEN_TOKEN, ANTHROPIC_API_KEY, and ORG_AUTOMATION_TOKEN (the existing token used by the other PR-opening workflows).

Idempotency

build_doc_catalog.py records each page's existing links, and the matcher folds them into the final set, so re-running does not duplicate links on a page. An item already represented on its target page with nothing better to add is dropped.