Pubmed Weekly Digest

PubMed Weekly Digest

A Claude Code skill: searches the past 7 days on PubMed across a configurable list of journals, refills missing abstracts via CrossRef, writes a TL;DR and a one-line Hot Take per article, and renders a Markdown digest to a local output directory. Committing or publishing the digest is up to you.

Configuration

Edit these values before running the skill in a new vault:

CROSSREF_MAILTO can also be supplied via an environment variable UNPAYWALL_EMAIL if you prefer not to bake it into the file. If neither is set, CrossRef calls still go out (some endpoints just rate-limit harder).

See README.md (same folder) for setup details and a step-by-step guide to swapping the journal list.

⚠️ ISO week calculation (read this first)

This skill runs at the start of a week and covers the just-ended week.

from datetime import date, timedelta

today = date.today()

# Search window: past 7 days (excluding today)
date_to   = today - timedelta(days=1)   # yesterday
date_from = today - timedelta(days=7)   # 7 days ago

# Week label: use the ISO week of the just-ended week, NOT today's week
target_date = today - timedelta(days=7)
iso_year, iso_week, _ = target_date.isocalendar()
week_label = f"{iso_year}-W{iso_week:02d}"   # e.g. 2026-W19

filepath = f"{OUTPUT_DIR}/{week_label}.md"

Example: today is 2026-05-11 (W20) → week_label = 2026-W19, search range 2026-05-04 to 2026-05-10.

Pitfall: do NOT use today.isocalendar() for the label — that mis-labels W19 content as W20.

Step 1 — PubMed search

Call mcp__PubMed__search_articles (datetype=edat) in parallel for each journal. The default list is six ID/haematology journals; replace freely:

date_from = date_from (YYYY/MM/DD), date_to = date_to (YYYY/MM/DD)

PubMed query caveats — these apply to every entry in the table:

• The MCP wrapper rejects wildcards (mycobacteri* → INVALID_PARAMETERS). Expand with OR instead: mycobacterium OR tuberculosis.
• The boolean-operator cap is 20. A single query like (a OR b OR ... OR u) fails; split into two parallel queries if needed.
• Add a topic filter with AND (term1 OR term2) to high-volume generals, e.g. N Engl J Med[Journal] AND (infection OR sepsis OR antimicrobial).

If search fails (MCP error or 0 articles for all journals): stop and report.

Step 2 — Fetch full metadata

Call mcp__PubMed__get_article_metadata in parallel batches (one batch per journal). Each batch is ≤ 20 PMIDs — the MCP silently truncates beyond that.

Step 2.5 — CrossRef abstract refill

Scan all articles. If abstract is empty (or "[Abstract not available]", which PubMed sometimes returns as a string) and a doi is present, query CrossRef:

• Endpoint: https://api.crossref.org/works/{urlencoded_doi}?mailto={CROSSREF_MAILTO}
• User-Agent: pubmed-weekly-digest/1.0
• Timeout: 15 seconds; cache by DOI within one run.
• If message.abstract is present in the response: strip JATS/HTML tags and collapse whitespace, then write back to the article's abstract field.
• If CrossRef returns no abstract, non-200, timeout, or parse failure: do NOT abort. Keep the article in a "still missing abstract" list.

Articles still missing abstracts after CrossRef:

• Interactive mode: list the affected articles (PMID, title, DOI) and ask the user to paste in abstracts manually, then continue. If the user says skip, leave TL;DR blank and note (no abstract) in the final output.
• Autonomous mode: leave TL;DR blank for those articles and list the PMIDs in the final report so the user can backfill on a future run.

Step 3 — Filter articles

Drop articles where article_types contains:

• "Erratum" or "Published Erratum"
• "Editorial" or "Comment" (no abstract → no meaningful annotation)

A section with zero surviving articles is fine — show a placeholder in Step 5.

Optional: per-journal type filtering

The default config includes an example for N Engl J Med where only Original Articles and Reviews are kept:

• article_types contains "Journal Article" but NOT "Case Reports", "Editorial", "Comment", "Letter", "News", "Biography", "Historical Article", "Portrait", "Interview", "Personal Narrative" → treat as Original Article
• article_types contains "Review" or "Systematic Review" → treat as Review Article
• Anything else from N Engl J Med → drop from this section

Adjust or remove this block for your own journal mix.

Step 4 — TL;DR and Hot Take annotation

For each surviving article:

• TL;DR: a 1–2 sentence English summary covering study design, key findings, and clinical implications. Keep drug names and pathogen names (bacterial / viral / fungal) in their canonical form; do not translate the article title.
• Hot Take: one short humorous English sentence per article. Pick a register that fits the content:
  • Positive — for hopeful findings, breakthroughs, good outcomes (excited / cheerful tone)
  • Snarky — for bad news, hard pathogens, depressing epidemiology, or old problems with no new answers (deadpan / self-aware tone, not cruel)

If abstract was never filled in, leave both fields blank.

Change the output language: this step is pure prompting. Replace "English" above with "Traditional Chinese", "Spanish", "Japanese", or whatever you prefer — Claude will translate the source abstracts on the fly. The rest of the skill is language-agnostic; only Step 4 needs to be edited.

Step 5 — Render Markdown

Write the digest to {OUTPUT_DIR}/{week_label}.md.

Read {SKILL_DIR}/template.md for the output scaffold.

Expand journal sections: for each journal in your Step 1 list, replace a ## {Journal N} stub with the real ## {Abbreviation} section using the article-block format in the template. Add extra sections if you have more journals than stubs; remove extras if fewer. Fill {journal_list} in the intro line with a comma-separated list of your journals.

Formatting rules:

• Authors: one author → full name; multiple → {first_author} et al.
• No DOI → PMID: {pmid} instead of the DOI line
• Section with no surviving articles → > No matching articles this week.
• Articles separated by ---

Step 6 — Save (optional commit)

The Markdown file lives at {OUTPUT_DIR}/{week_label}.md. Stop here unless the user explicitly asks to commit/push — this fork intentionally does not assume a publish target.

If the user wants the file committed to their vault:

git add {OUTPUT_DIR}/{week_label}.md
git commit -m "weekly journal digest {week_label}"

If they want to push to a public site repo, that's their choice — add the relevant git remote / gh pr / API call as a follow-up step.

Error Handling