brycewang-stanford/literature-search

Literature Search

Overview

This skill searches academic literature with strict web verification of every result. The primary anti-pattern it prevents is hallucination of citations — plausible-sounding papers that do not exist. Every paper returned must be fetched from the web and confirmed to be real before it enters the output.

When to Use

• "What does the literature say about X?"
• "Find papers on Y"
• "Who wrote the seminal paper on Z?"
• Building a reference list for a new project
• Verifying that a specific paper exists
• Checking the correct citation details for a paper the user remembers partially

Modes

This skill operates in two modes. The caller specifies which.

Gap-check mode

Used during brainstorm to verify a research gap exists. Scope: 1-2 targeted queries, 5-8 results maximum, no bibliography file output. Web-verify every result (the anti-hallucination rule still applies). Output the standard markdown table but do not populate references.bib.

Full mode

Used during execute-plan's Literature phase. Run all Mandatory Steps below. Curate 15-30 references, bias toward target journals, populate references.bib via citation-management, produce literature notes. This is the default when no mode is specified.

Mandatory Steps

1. Identify the research field from context. Resolve CLAUDE.superpapers.md by reading it from the current working directory, or walking up parent directories until found. If the file contains a field entry, use it. Otherwise infer from the project abstract or ask the user directly. The field determines which databases to prioritize.

2. Choose databases based on the field:

   • Core (any field): Google Scholar, Web of Science, JSTOR, Semantic Scholar
   • Economics and finance: SSRN, NBER, RePEc, ScienceDirect
   • Health and medicine: PubMed, Cochrane Library
   • Political science and sociology: JSTOR, SAGE Journals
   • Physical and computer sciences: arXiv, ACM Digital Library, IEEE Xplore

3. Search with specific queries, not broad terms. Include methodology keywords when relevant. For example, prefer "difference-in-differences" minimum wage employment over minimum wage.

4. For every candidate paper, web-fetch the landing page. Confirm title, authors, year, and DOI from the authoritative source. If the paper cannot be fetched and verified, mark it as [unverified] and exclude it from the final list.

5. Output a markdown table with columns: Authors (Year), Title, Venue, DOI, Relevance. Keep the relevance column to a single line of substantive justification.

6. Curate, do not dump. Prioritize seminal works, recent papers (last five years), highly cited results, and papers that directly address the user's question. Aim for 8 to 15 results per query unless the user asks for more.

7. Bias toward the user's target journals. Use the target journal list resolved from CLAUDE.superpapers.md in step 1, or ask the user if the file was absent or the field was unset. Ensure a substantial share of cited references come from those journals or closely related ones in the same field tier. Prioritize very recent publications (last three to five years) in those outlets — reviewers expect to see that the authors know the journal's recent conversation. If a search returns few results from the target journals, widen to journals of similar scope and rank, but flag the gap to the user.

Output Format

| Authors (Year) | Title | Venue | DOI | Relevance |
|---|---|---|---|---|
| Card & Krueger (1994) | Minimum Wages and Employment | American Economic Review | 10.xxxx/xxxxx | Seminal DiD on minimum wage; identifies employment effects |
| Dube et al. (2010) | Minimum Wage Effects Across State Borders | Review of Economics and Statistics | 10.xxxx/xxxxx | Border-pair design, finds no disemployment effect |

Replace example DOIs with real verified values. Never leave placeholder DOIs in the final output.

Anti-Patterns

• Primary: fabricating a paper that does not exist. Inventing "Smith (2021), Some Plausible Title, Journal of Plausible Results" without verification.
• Citing from memory without web fetch
• Listing 30 papers without curation of relevance
• Including working papers alongside published versions without distinction
• Confusing a journal article with a book chapter or conference paper
• Listing papers outside the user's field
• Returning results without DOIs or stable URLs
• Trusting a paper's existence because a well-known author "probably wrote it"
• Ignoring the target journal list when curating results — the bibliography should reflect the journal's recent conversation

Verification Before Completion

• [ ] Every paper in the final list was fetched via web in this session
• [ ] Every paper has a verified DOI or stable URL
• [ ] No [unverified] entries in the final table
• [ ] Field-appropriate databases were used
• [ ] Relevance column is populated with substantive reasoning
• [ ] Results are curated, not dumped
• [ ] Working papers distinguished from published versions
• [ ] Substantial share of references come from the user's target journals or closely related outlets
• [ ] Recent papers (last 3-5 years) from target journals are well represented