Literature Search Skill

Every reference you produce must come from a live database result — never generate a citation from memory alone, because a recalled citation can look real and not exist.

Search Tools: MCP (Primary) + E-utilities (Fallback)

Primary: MCP Tools (Claude.ai Remote)

Fallback: NCBI E-utilities (Direct API via Bash)

If any mcp__claude_ai_PubMed__* call returns an error containing "terminated", "not found", "not available", or "not connected", switch ALL subsequent PubMed calls in this session to the bundled E-utilities scripts. Do not retry MCP after a disconnect — it will not recover within the same conversation.

EUTILS="${CLAUDE_SKILL_DIR}/references/pubmed_eutils.sh"
PARSER="${CLAUDE_SKILL_DIR}/references/parse_pubmed.py"

# Search PubMed (returns PMIDs)
bash "$EUTILS" search "diagnostic test accuracy meta-analysis radiology" 20 \
  | python3 "$PARSER" esearch

# Get article summaries as markdown table
bash "$EUTILS" fetch_json "16168343,16085191,31462531" \
  | python3 "$PARSER" esummary

# Get detailed metadata
bash "$EUTILS" fetch "16168343" \
  | python3 "$PARSER" efetch

# Generate BibTeX entries
bash "$EUTILS" fetch "16168343,16085191" \
  | python3 "$PARSER" bibtex

# Verify a citation by exact title
bash "$EUTILS" cite_lookup "Bivariate analysis of sensitivity and specificity" \
  | python3 "$PARSER" esearch

# Find related articles for a PMID
bash "$EUTILS" related "16168343" 10 \
  | python3 "$PARSER" esummary

The script sleeps 350 ms between calls (NCBI allows 3 requests/s without an API key, 10/s with NCBI_API_KEY); keep batch calls sequential.

Workflow

Phase 1: Search Strategy

1. Get the research topic, question, or manuscript section that needs references.
2. Build the query from the key concepts (Population, Intervention/Exposure, Comparison, Outcome),

with MeSH terms for PubMed: (concept1 OR synonym1) AND (concept2 OR synonym2).

1. Set scope: date range (default: last 10 years unless the user specifies), article types, and

language (default: English).

1. Present the Boolean query, databases, and filters to the user.

Gate: Wait for user approval before running searches.

Phase 2: Execute Search

1. Search PubMed (search_articles, Boolean query), Semantic Scholar (semanticSearch, natural

language query), and bioRxiv/medRxiv (search_preprints) when preprints are relevant.

1. Deduplicate across databases by DOI or title similarity.
2. Present the results in this table, and write the same rows to references/search_results.tsv:

| # | Title | Authors (first + last) | Year | Journal | PMID/DOI | Relevance |
|---|-------|----------------------|------|---------|----------|-----------|
| 1 | ...   | Kim J, ... Lee S     | 2024 | Radiology | 12345678 | High      |

1. Ask the user to select which papers to include.

#### Record what the source said existed, not only what you downloaded

A wrong haul raises no error, and a PRISMA flow built on a wrong number is fiction that nothing downstream contradicts. Check both of these:

• A count that equals a page cap exactly. Every source reports a total:

esearchresult.count, opensearch:totalResults, meta.count. Record `api_total` beside `downloaded`, and fail loudly when `downloaded < api_total`, or when `downloaded` equals a page or loop cap exactly (e.g. a loop's own if start >= 2000: break reporting 2,000 records). Print TRUNCATED and refuse to write the search record.

• A boolean that was never applied. OpenAlex's search= is a relevance-ranked free-text

parameter that silently ignores AND/OR; filter=title_and_abstract.search: honours them. Run the query once more with one mandatory clause negated. If the hit count does not drop, the boolean is being ignored — the engine is ranking, not filtering.

PubMed via E-utilities is the one place where the naive pattern is safe. Everywhere else, do both.

#### A DOI in a screening row is not necessarily that row's DOI

When the pipeline filled a doi column (matched against Crossref by title similarity) rather than receiving it with the record, a wrong match is a valid, resolvable DOI for a different paper. Resolve it and read the title back before any decision rests on it:

python3 "${CLAUDE_SKILL_DIR}/scripts/check_doi_record_match.py" --table 2_Screening/round3.tsv \
        --email <contact> --json qc/doi_record_match.json

DOI_NOT_THIS_RECORD is a DOI that resolves to another paper; DOI_IS_CONTAINER resolves to an issue, supplement or proceedings rather than an article; DOI_IS_UPDATE_NOTICE resolves to a correction, erratum or retraction notice (Crossref update-to) rather than to the article it updates; DOI_UNRESOLVED is reported rather than dropped. This runs at screening, where a wrong DOI is still cheap; /verify-refs audits a finished reference list.

Phase 2.5: Citation Searching (Snowballing)

Optional; recommended for systematic reviews and thorough background work (PRISMA item 7, "records identified through citation searching"). Expand a seed set along the citation graph with the Semantic Scholar Graph API helper:

python3 "${CLAUDE_SKILL_DIR}/references/snowball.py" \
  --seed DOI:10.1000/synthetic.example,PMID:00000000 \
  --direction all \
  --pool references/library.bib \
  --out references/library.bib

• Directions: backward (references the seeds cite), forward (papers citing the seeds),

similar (S2 recommendations), or all (default). Dedup is against the --pool and within the harvested set, by DOI and normalized title.

• Trust flag: candidates are written verified=false + verified_by=semantic_scholar. Run

/verify-refs (or Phase 4 verification) on each before citing it.

• Output contract: appends to references/library.bib only; the script hard-refuses

manuscript/_src/refs.bib.

• PRISMA line: the script prints `Records identified through citation searching (snowballing):

N raw (backward=…, forward=…, similar=…); after dedup against existing pool: M new candidates.` Record M in the PRISMA flow's citation-searching box.

• Incomplete runs: when any seed/direction fetch fails, or the source says more records exist

past --limit (a next page), the script prints a FAILED or TRUNCATED line per seed/direction, ends the PRISMA line with INCOMPLETE, and exits 1. Those counts are a lower bound. Do not record them; re-run (or raise --limit) until the run exits 0.

Phase 3: Deep Read

For each selected paper, retrieve full metadata (get_article_metadata for PubMed, get_preprint for bioRxiv) and extract the study design, sample size/dataset, key methods, primary findings (with specific numbers), and the authors' stated limitations. For several papers, present a literature matrix for review:

| Paper | Design | N | Key Finding | Limitation | Relevance to Our Study |
|-------|--------|---|-------------|------------|----------------------|

Phase 4: Citation Management

#### Verification

1. NEVER fabricate a DOI or PMID. If you cannot find one, mark the reference

[UNVERIFIED - NEEDS MANUAL CHECK].

1. Cross-check every reference against the API result: first and last author, publication year,

journal, article title (exact, not paraphrased), and volume/pages when available. Flag each field that does not match.

1. Confirm each DOI resolves with WebFetch https://api.crossref.org/works/{DOI} (on CrossRef

errors, follow Error Handling).

#### BibTeX Generation

Generate an entry for every reference, verified or not, with an explicit verified flag:

@article{FirstAuthorLastName_Year_ShortKey,
  author    = {Last1, First1 and Last2, First2 and Last3, First3},
  title     = {Full Title As Retrieved From Database},
  journal   = {Journal Name},
  year      = {2024},
  volume    = {310},
  number    = {2},
  pages     = {e234567},
  doi       = {10.1001/jama.2024.12345},
  pmid      = {12345678},
  verified  = {true},
  verified_by = {pubmed+crossref},
  verified_on = {2026-04-24},
}

verified_by lists the confirming sources (pubmed, crossref, semantic_scholar, or a combination); verified_on is the ISO date of the most recent successful verification. No downstream script reads these fields — /verify-refs re-checks every entry — so they record trust rather than grant it.

BibTeX key convention: FirstAuthorLastName_Year_OneWord (e.g., Kim_2024_Validation). The key is provisional: Better BibTeX assigns the citable key when /lit-sync imports the entry into Zotero.

#### Output

1. Append the entries to references/library.bib (do not overwrite) — the candidate pool

/lit-sync imports into Zotero. NEVER write to manuscript/_src/refs.bib, because /lit-sync (via Better BibTeX) is its sole writer.

1. Print a summary with verification status:

Verified:    12 references (verified=true)
Unverified:   1 reference  (verified=false) [NEEDS MANUAL CHECK]
Total:       13 references

Phase 4b: Zotero Library Integration

Importing into Zotero belongs to /lit-sync, which owns Zotero writes and references/zotero_collection.json: hand it references/library.bib. If a Zotero MCP server is connected, you may read from it here — zotero_search_items (by DOI) to mark candidates already in the library, zotero_get_annotations to reference the user's prior reading notes.

Phase 5: Full-Text Retrieval

Delegate to /fulltext-retrieval, the single home of the open-access cascade; do not re-implement OA fetching here. Pass the verified candidate DOIs from references/library.bib:

ENGINE="${CLAUDE_SKILL_DIR}/../fulltext-retrieval/fetch_oa.py"
# extract DOI + Title (and PMID/FirstAuthor when available) → worklist.tsv
python3 "$ENGINE" worklist.tsv -o pdfs/ -e <contact-email> --report pdfs/retrieval_report.json

Put the verified bibliographic title in the worklist rather than a DOI-only list. Keep source_identity and file_sha256 from the retrieval report with the record. Download success and title agreement alone do not verify the PDF: inspect conflicts, unresolved/unavailable evidence, and files whose hashes have changed before citing them. Missing identity fields in older reports mean unassessed; even consistent is advisory front-matter corroboration, not verification of the paper's claims.

For Zotero-resident PDFs and proxy-aware retrieval, use /lit-sync Phase 2.7. For DOIs in pdfs/manual_needed.txt, use only institutional access (your library's own subscriptions, proxy or VPN), interlibrary loan, or the corresponding author. Never bypass paywalls or publisher access controls, and do not configure unauthorized PDF mirrors.

Phase 6: Gap Analysis

When called during manuscript writing, extract the manuscript's inline citations, compare them with the search results, and report specific gaps: key papers not cited, outdated references with newer versions, and missing methodological references (statistical methods, reporting guidelines).

Specialized Search Modes

Mode: Manuscript Paper Reference Pool

Supplies a manuscript's reference pool — typically invoked by /write-paper Step 7.3c (or /self-review Phase 2.5c-2) when the reference-adequacy gate finds the draft under target or a named method uncited; usable directly for an original-research bibliography.

For an original-research article, return 25–40 verified candidates, not the ~10 a quick search settles on. If the field is genuinely sparse, say so explicitly rather than returning a thin list silently. Respect a narrower journal reference cap or user scope when one is given.

Cover six candidate categories:

1. Background / disease burden / clinical context — why the question matters.
2. Gap-defining prior studies — the work the manuscript extends or contradicts.
3. Comparator / comparable-design cohorts — studies the Results will be measured against.
4. Methods / statistical canonical sources — the originating reference for every named method,

model, score, equation, or diagnostic criterion (e.g. competing-risk model, multiple imputation, E-value, eGFR equation, concordance statistic). This category clears Methods named-method gaps.

1. Reporting-guideline sources — STROBE, TRIPOD(+AI), CONSORT, PRISMA(-DTA), STARD, etc.
2. Interpretation / mechanism / limitation support — grounds Discussion claims.

For each candidate, report PMID/DOI, verification status, candidate category, the target manuscript section, and a one-line why it is needed. Entries go through Phase 4 into references/library.bib only. This mode produces candidates: the user decides inclusion, and it does not insert references into the manuscript bib.

Mode: Crowding Check

Run before a study is designed. A background search ("what has been written about this topic") leaves the trap open; ask four narrower questions instead:

Give each its own verdict: a design can be original on one axis and fully occupied on another, and collapsing the four into one answer hides that. The journal row is not vanity: a design once matched an existing paper on frame, coding axis and journal — its own first choice. Most of the time this mode narrows a claim rather than ending a project (e.g. from "nobody has looked at this" to "nobody has decomposed it by provenance"), and the narrowed claim survives review.

Search the way a competitor would: the exact frame, the exact measure, and the journal's own site, not only the topic. Report the four verdicts and the papers behind each, then let the user decide.

Mode: Systematic Search

For systematic reviews or comprehensive literature sections:

1. Document the full search strategy (PRISMA-compliant).
2. Record: database, date of search, query string, number of results.
3. Track inclusion/exclusion at each screening step.
4. Output a PRISMA flow diagram data summary.

Mode: Quick Cite

For a single reference the user describes ("that 2023 paper by Smith about AI in chest X-ray"): search PubMed and Semantic Scholar with the details, present the top 3 candidates, and generate the BibTeX entry for the one the user confirms.

Mode: Related Papers

From a PMID or DOI, get related papers with find_related_articles plus Semantic Scholar citation-based recommendations, ranked by relevance. For a structured, dedup-aware, PRISMA-countable expansion (backward + forward + similar), use Phase 2.5: Citation Searching with references/snowball.py instead.

Mode: Embase Browser Automation

Embase has no public API. Read ${CLAUDE_SKILL_DIR}/references/embase_browser.md when the search must include Embase — it has the Chrome-automation export steps, the CSV row format, and the PubMed → Embase query translation.

Error Handling

• If a search returns 0 results, broaden the query (remove one concept or use broader MeSH terms)

and retry.

• CrossRef HTTP errors (token-saving rules):
• 403 (rate-limited): Do NOT retry. Skip CrossRef → verify via PubMed title search instead.
• 303 (redirect): Follow the redirect if possible. If not, skip CrossRef → PubMed fallback.
• After the first CrossRef 403/303 in a session, skip CrossRef for ALL remaining references and

go directly to PubMed title verification, to avoid N×retry token waste.

• Do not print raw error messages ("Request failed with status code 403."). Report one summary

line at the end: CrossRef unavailable for {N} references (rate-limited). Verified via PubMed instead.

• If a DOI does not resolve via CrossRef (after the rules above), search PubMed by title to confirm

the reference exists.

• If a reference cannot be verified by any method, state: "This reference could not be verified.

Please check manually before submission." Never silently include an unverified reference.

Known limits

• check_doi_record_match.py compares titles at a similarity threshold and reads structured

Crossref fields (type, update-to). A study protocol, or part 2 of a multi-part paper, whose title differs from the row's by a word or a number is not separated from the paper the row describes; no structured field marks it. Treat a silent run as "no mismatch detected", not as proof that every DOI is right.