Use Scholar Feed as a skill in Claude

What's a skill?

A skill is a single markdown file that teaches Claude how to compose multiple MCP tools into a research workflow. Instead of you remembering “run search_papers, then find papers that cite the anchor and engage critically, then fetch_fulltext on the top 3 results, then synthesize,” you type /find-contradictions and Claude follows the recipe.

Skills live in the public Scholar Feed GitHub repo (see skills/). Claude Code discovers them automatically via the Anthropic Skills native distribution channel. No separate install — once the Scholar Feed MCP is connected, the skills are available.

One-time install

If you already have the Scholar Feed MCP installed, skip this. If not:

claude mcp add scholar-feed -- npx -y scholar-feed-mcp

Anonymous tier (100 requests/day) requires no API key. For higher limits see the developers page.

/find-contradictions

First shipped skill

Surface qualitative disagreements between research papers — competing claims, conflicting methods, opposing findings, or contested interpretations — by composing semantic search, citation context, and LLM judgment over the Scholar Feed corpus.

Trigger phrases Claude recognizes

• “which papers disagree about X”
• “what are the contested claims in field Y”
• “is there a debate about method Z”
• “find me critiques of paper W”
• “where does the literature conflict on topic A”

Example session

Use /find-contradictions to find papers that disagree about whether scaling laws hold for sparse models.

What it composes

• search_papers — discover candidate papers on the topic, with multiple query reformulations to surface papers that broad phrasing misses
• search_papers(scope_to_citations_of=ANCHOR, q=axis) — find papers that cite the anchor AND engage critically with the disagreement axis
• get_paper and fetch_fulltext — verify the disagreement is real, not topical overlap or terminology drift
• LLM judgment in Claude — distinguish substantive disagreement from passing mention

Output format

For each verified disagreement, the skill produces a structured tuple:

What it does NOT do

/find-contradictions is qualitative only. It does NOT surface numerical comparisons like “Paper A reports 73.2 on MTEB, Paper B reports 71.5.”

The corpus does not carry reliable head-to-head benchmark numbers. Any numerical comparison built on top would be silently stale or wrong — the same failure mode that retired the leaderboard tool in v2.0.0. The skill instead surfaces claim-level, method-level, finding-level, and interpretation-level disagreements that are visible in paper summaries, abstracts, and citation context.

If the literature is consistent on the axis you asked about, the skill says so. Negative results are valid output — the skill will not manufacture controversy that does not exist in the papers.

Paper-writing skills

A bundle of three skills for the paper-writing workflow: /related-work-draft writes a grounded related-work section with inline citations and a matching .bib block, /bib-audit reviews your bibliography against the live corpus before you submit, and /reviewer-triage is the Area Chair workflow — topical fit plus co-authorship conflict detection in one prompt.

/related-work-draft

Draft a grounded related-work section

Retrieves relevant prior work from the Scholar Feed corpus, fulltext-grounds the key claims, and emits a draft related-work section as flowing prose with inline \cite{...} markers plus a matching .bib block ready to paste into Overleaf.

Trigger phrases Claude recognizes

• “draft a related work section about X”
• “write a literature review on Y”
• “help me write related work for my paper on Z”
• “give me a related-work paragraph on topic A with citations”
• “synthesize prior work on C with bibtex”

Example session

Use /related-work-draft to draft a related-work section on retrieval-augmented LLMs for scientific reasoning. Output prose with \cite{...} markers and the .bib block.

What it composes

• search_papers — retrieve candidate papers per sub-axis, with multiple query reformulations to surface phrasing-sensitive papers
• embed_text — optional HyDE composition when the natural-language query is too vague
• search_papers(scope_to_citations_of=ANCHOR, q=axis) — find follow-up work that cited and engaged with a foundational anchor
• fetch_fulltext — ground every specific factual claim in the paper's own text before citing it
• get_paper(arxiv_ids=[...], format='bibtex') — batch-fetch BibTeX in one call

Scope

No hallucinated citations. Every \cite{...} key produced by the skill MUST correspond to a paper retrieved from the corpus. If the corpus does not surface a paper for a claim, the draft either omits the claim or marks it [citation needed] — never invents a citation.

/bib-audit

Pre-submission bibliography review

Batch-checks every entry in your .bib against the Scholar Feed corpus and emits a 4-section audit: confirmed-present entries, drift entries (in corpus but metadata differs), not-found entries (could not be matched to an arxiv_id — flagged for verification), and optional missing-recommendations (canonical work the corpus surfaces for your topic that's not in your .bib).

Trigger phrases Claude recognizes

• “audit my .bib file”
• “check my bibliography against the corpus”
• “review my .bib before I submit”
• “are there better citations I'm missing for my paper on X”
• “did I get the author list right for [paper]”
• “is my arxiv version stale”

Example session

Use /bib-audit on paper/references.bib. Flag drift in author lists, years, and titles. Skip missing-recommendations.

What it composes

• get_paper(arxiv_ids=[...]) — ONE batch call with every extracted arxiv_id to fetch corpus metadata
• search_papers — title-based discovery for .bib entries with no arxiv_id
• fetch_fulltext — ground any suspicious drift flag (e.g. arxiv-v1 vs v3 title change)
• LLM judgment — diff metadata, distinguish substantive drift from cosmetic, format the 4-section audit report

Scope

“Not in corpus” does not mean “does not exist.” The Scholar Feed corpus is ~600k arxiv papers, heavily concentrated in cs.LG / cs.AI / cs.CL. NeurIPS proceedings without arxiv preprints, books, journal-only publications, and pre-2018 papers are commonly outside the corpus. The skill flags Not-found entries for your verification with that caveat explicit; it never recommends deletion.

/reviewer-triage

Area Chair reviewer-pool triage

Takes a paper submission and emits a 3-section triage report: Recommended reviewers (topically qualified, no co-authorship edge to submission authors in the configurable window), Conflict-flagged (topically qualified with co-authorship edge detected), and Suggested-but-verify (qualified, no co-authorship in corpus but you should layer your venue's full CoI policy on top).

Trigger phrases Claude recognizes

• “find reviewers for this paper”
• “suggest reviewers for submission X”
• “who could review this without a conflict of interest”
• “AC workflow — triage reviewers for paper Y”
• “who's worked on topic Z and isn't a co-author of the submitting team”

Example session

Use /reviewer-triage for arxiv:2509.25085. Use a 3-year CoI window. Surface 10 candidates per section.

What it composes

• search_papers — find topically qualified candidate reviewers via recent work on the submission's topic
• co_author_graph(author_ids=[submission + candidates], window_years=N) — ONE call surfaces every co-authorship edge between submission authors and candidates (Phase 111 primitive)
• get_paper — optional shortlist metadata to sanity-check fit
• LLM judgment — rank by topical depth, format the 3-section triage report, attach co-authored-paper evidence to every conflict flag

Scope

Co-authorship is ONE conflict axis. The skill detects co-authorship on papers in the corpus within the configurable window_years (default 10). It does NOT detect institutional affiliation conflicts, advisor-advisee ties beyond co-authored papers, funding-source conflicts, or personal-relationship conflicts. You must layer your venue's full CoI policy on top. “Recommended” from this skill means “no co-authorship edge found in corpus within the window” — not “no CoI exists.”

More skills

Skill files live in the public Scholar Feed GitHub repo: skills/. Claude Code discovers them automatically once the MCP is connected.