OpenAlex Skill

> FEATURED — Teaches an agent how to use OpenAlex (the open scholarly graph) correctly: discover entities, query works with filters, read results, and cite accurately. No API key required. Set OPENALEX_MAILTO for the polite pool.

This skill pairs with the OpenAlex MCP server (see ../mcp/), which provides the 6 callable tools. The skill provides the know-how. Use imperative voice; do what each step says.

1. Name

openalex — Open scholarly metadata: works, authors, institutions, sources, topics, concepts, publishers, funders.

2. Purpose

Answer research questions using authoritative bibliographic data: find papers, authors, citations, open-access status, and bibliometric trends — and cite them precisely. OpenAlex is free and open.

3. When to use OpenAlex

Use OpenAlex when the task involves:

• Scholarly works (papers, preprints, datasets, books) and their metadata.
• Authors and their output, affiliations, and citation counts.
• Citations and impact (cited_by_count, FWCI).
• Open-access status and finding free full-text links.
• Bibliometrics / trends: counts by year, institution, topic, OA status.
• Institutions, journals/sources, topics, concepts, publishers, funders.

It is free — prefer it for any academic-metadata need.

4. When NOT to use OpenAlex

• Full-text PDFs / reading the paper body → OpenAlex gives metadata + open_access.oa_url; follow that URL to the file. OpenAlex does not serve full text.
• General/non-academic web information → use a web search API, not OpenAlex.
• Paywalled full text → OpenAlex can tell you if/where an OA copy exists, but cannot bypass paywalls.

5. Environment

• No API key. No required environment variables.
• Recommended: set OPENALEX_MAILTO=you@example.com to join the polite pool (faster, fewer 429s). Not a secret.
• Optional: OPENALEX_API_BASE_URL, OPENALEX_TIMEOUT_MS (30000), OPENALEX_MAX_RETRIES (3), LOG_LEVEL.

6. Operations (the 6 tools + generic)

7. Discovery workflow

1. Start from human input (a name, title, keyword).
2. Resolve to an entity ID with openalex_search or openalex_request → autocomplete/{entity}.
3. Verify you picked the right entity (check display_name, affiliation, works_count).
4. Note the ID prefix → entity type:

Entity types: works, authors, sources, institutions, topics, concepts, publishers, funders, keywords.

8. Query workflow

Build a filter (comma-separated, ANDed) and pick a sort:

• Sort by impact: cited_by_count:desc. Sort by recency: publication_date:desc.
• per-page ≤ 200.
• For deep traversal, use cursor (cursor=* then meta.next_cursor), not high page numbers.

9. Reading results

• meta.count = total matches (not the number returned).
• results = the current page only.
• group_by = [{key, key_display_name, count}] for aggregations.
• Abstract: works carry abstract_inverted_index (a {word: [positions]} map), not plain text. Reconstruct by placing each word at its positions and joining in order.
• Full text: follow open_access.oa_url for the free PDF/HTML.

10. Citation rules

Cite every claim with: title, authors, year, DOI, and the OpenAlex ID + URL https://openalex.org/.

<Authors> (<year>). <Title>. <Source>. DOI: <doi>. OpenAlex: https://openalex.org/<WID>

The OpenAlex URL is mandatory for traceability, in addition to the DOI.

11. Freshness

OpenAlex data updates frequently (new works, citation counts, affiliations). Counts you report are point-in-time. When precision matters, note the access date and that figures may change.

12. Integrity

• Report only what the API returns. Never invent papers, authors, DOIs, or citation counts.
• If results are empty, say so and broaden — do not fabricate to satisfy a requested count.
• Keep totals (meta.count) distinct from listed results.

13. Error handling

14. Cost / etiquette

• Free. Be polite: set OPENALEX_MAILTO.
• Cache resolved IDs and stable records.
• Avoid huge unfiltered scans. Always filter first.
• Use cursor, not high page numbers (page is capped ~10000 results).

15. Security

• No secrets to manage. OPENALEX_MAILTO is not sensitive but keep configs clean.
• Read-only API; outbound HTTPS only. Keep logs on stderr; protocol on stdout.

16. Agent checklist

• [ ] Resolved names to IDs (and verified the right entity)?
• [ ] Built a filter instead of scanning everything?
• [ ] Chose an appropriate sort?
• [ ] Used cursor for deep paging?
• [ ] Read meta.count vs results correctly?
• [ ] Reconstructed abstracts from the inverted index if needed?
• [ ] Cited title + authors + year + DOI + OpenAlex ID/URL?
• [ ] Set OPENALEX_MAILTO to avoid 429?
• [ ] Reported only real, returned data?

17. Example workflows

• Literature review: resolve topic → openalex_works (topic + year + is_oa, sort by citations) → openalex_get top work → author/institution profiles → cited summary. See recipes/literature-search.md.
• Author profile: resolve author → openalex_get author → openalex_works filtered by authorships.author.id → top works + metrics. See recipes/author-profile.md.
• Trend by year: openalex_group_by on publication_year with a topic/OA filter. See recipes/citation-trends.md.

18. Common mistakes

• Using per_page on the wire instead of per-page (hyphen) in openalex_request.
• Deep-paging with high page numbers (capped ~10000 results) instead of cursor.
• Treating abstract_inverted_index as plain text.
• Reporting meta.count as the number of items returned.
• Forgetting the OpenAlex ID/URL in citations.
• Skipping OPENALEX_MAILTO and hitting 429.

19. Maintenance

• Re-resolve IDs periodically; entities can merge/change.
• Re-check filter keys and limits against when behavior changes.
• Update cached records given frequent data refreshes.

> Verification needed: confirm filter keys, limits, and field names with .