Literature Review skill for Claude Code

Done well means: a reader can rerun the search from the documented strings and dates, every included source exists and says what the review claims it says, the synthesis is organized by ideas (not by paper), and the gaps are argued from the matrix rather than asserted.

Hard rules

• Never cite a paper you have not located. Every entry needs a resolvable DOI, arXiv ID, PubMed ID, ISBN, or stable URL that you actually opened or that the user supplied. If a search tool is unavailable, say so and produce the search plan and matrix template instead of a reference list from memory.
• Never invent page numbers, sample sizes, effect sizes, or quotes. Mark anything not read in full as abstract-only.
• Keep the review’s claims proportional to the evidence: one study “reports”, several consistent studies “indicate”, a pre-registered meta-analysis “shows”.
• A narrative review is not a systematic review. If the user needs exhaustive, protocol-driven, bias-assessed evidence (clinical, policy, meta-analysis), switch to the systematic-reviews workflow.

Choose the review type first

Need	Type	Search depth	Output
Background or related-work section	Narrative	2-3 databases + citation chasing	1-4 pages, thematic
Map what exists on a broad topic	Scoping	Wide, documented, no quality filter	Evidence map + categories
Answer a precise effect question	Systematic	Exhaustive, protocol, dual screening	PRISMA flow, bias tables
Quick answer for a decision	Rapid	1-2 databases, single screener, date-limited	Brief with stated limits
Learn a field fast	Reading list	Seminal + recent reviews	10-20 annotated items

State the chosen type and its limits at the top of the deliverable.

Workflow

- [ ] 1. Frame the question and scope
- [ ] 2. Build and log the search
- [ ] 3. Screen against written criteria
- [ ] 4. Chase citations forward and backward
- [ ] 5. Extract into the synthesis matrix
- [ ] 6. Synthesize by theme; argue gaps
- [ ] 7. Verify every reference
- [ ] 8. Write up with the search appendix

1. Frame the question

Turn the topic into a question with explicit components. Use the framework that fits:

• PICO (interventions): Population, Intervention, Comparison, Outcome.
• PEO (exposures/qualitative): Population, Exposure, Outcome.
• SPIDER (qualitative/mixed): Sample, Phenomenon of Interest, Design, Evaluation, Research type.
• Concept-Context (CS/engineering): technique + task/domain + evaluation setting.

Write down: years covered and why, languages, publication types (peer-reviewed, preprints, grey literature), and the disciplines whose databases matter. Ask the user when the scope is ambiguous; a review of “LLM hallucination” since 2020 and since 2023 are different projects.

2. Build and log the search

1. List 2-4 concepts from the question. For each, collect synonyms, spelling variants, acronyms, and controlled-vocabulary terms (MeSH, ACM CCS, ERIC thesaurus).
2. OR within a concept, AND across concepts. Use truncation and phrase quotes.
3. Test against 3-5 known “gold” papers the user or a recent review identifies. If the string misses one, find out which term it lacks and add it.
4. Log every run: database, date, exact string, filters, hit count.

("large language model*" OR LLM* OR "foundation model*")
AND (hallucinat* OR "factual consisten*" OR confabulat*)
AND (detect* OR mitigat* OR evaluat*)

Database choice by field:

Field	Primary sources
Biomedicine / health	PubMed/MEDLINE, Embase, Cochrane Library, CINAHL
Computer science / AI	ACM DL, IEEE Xplore, DBLP, arXiv, Semantic Scholar, OpenReview
Social science / education	Scopus, Web of Science, ERIC, PsycINFO, SSRN
Economics / policy	RePEc/IDEAS, NBER, SSRN, Google Scholar, government/IGO repositories
Cross-disciplinary	Scopus, Web of Science, OpenAlex, Google Scholar (supplementary only)

Google Scholar results are not reproducible (ranking changes, 1,000-result cap, no full Boolean support): use it for citation chasing and gap filling, never as the only source. For field-specific syntax (PubMed tags, Scopus TITLE-ABS-KEY, arXiv API queries, OpenAlex filters), read references/search-syntax.md.

3. Screen against written criteria

Write inclusion/exclusion criteria before screening and apply them in two passes: title/abstract, then full text. Record the reason for each full-text exclusion.

Include: empirical studies; evaluates a hallucination detection or mitigation method;
English; 2020-2025; peer-reviewed or arXiv with >= 1 version and code or data.
Exclude: opinion pieces; image-only models; no quantitative evaluation; duplicate
versions (keep the latest/peer-reviewed one).

• Deduplicate before screening (DOI first, then normalized title + year). The citation-management skill has a dedupe script.
• When uncertain at title/abstract stage, include and decide at full text.
• Keep counts for each stage; even a narrative review should report them in one line.

4. Chase citations

• Backward: scan reference lists of included papers and recent reviews.
• Forward: “cited by” in Google Scholar, Semantic Scholar, Scopus, or OpenAlex (cites: filter) for seminal papers.
• Stop when two consecutive rounds add no new includable papers (saturation) and say so.

5. Extract into a synthesis matrix

One row per study; columns chosen from the question. Fill from the full text when possible, and flag abstract-only rows.

| ID | Citation (verified) | Year | Setting/data | Method | N / scale | Key finding | Limits | Theme(s) |
|----|---------------------|------|--------------|--------|-----------|-------------|--------|----------|
| S01 | Doe et al., doi:10.xxxx/yyyy | 2023 | QA, 3 datasets | Self-consistency sampling | 5 models | -18% hallucination rate vs. greedy | English only; no human eval | detection, sampling |

6. Synthesize and argue gaps

• Organize by theme, mechanism, or chronology of ideas, not one paragraph per paper.
• For each theme: what is established (convergent evidence), what is contested (conflicting results and the likely reason: data, metric, population), what is missing.
• Derive gaps from empty or thin cells in the matrix: understudied populations, settings, methods, outcomes, or time periods. Classify each gap:

Gap type	Signal in the matrix
Evidence gap	Contradictory findings with no resolving study
Methodological gap	All studies use the same design, metric, or benchmark
Population/context gap	Settings or groups absent from the “Setting” column
Empirical gap	Claim repeated in intros but never tested
Theoretical gap	Findings with no explanatory model connecting them

• Note publication-bias risk: field dominated by positive results, benchmark-driven reporting, or industry-only datasets.

7. Verify every reference

For each reference, confirm in a primary index (doi.org, Crossref, PubMed, arXiv, publisher page):

• Title, authors, year, and venue match.
• The cited claim appears in the paper (not just the abstract of a different version).
• Retraction status (Crossref metadata includes Retraction Watch data; check the publisher page and PubMed for retraction notices).
• Preprint vs. published version: cite the version you read; note if a later version changed conclusions.

curl -sL -H "Accept: application/x-bibtex" https://doi.org/10.1038/s41586-020-2649-2
curl -s "https://api.crossref.org/works/10.1038/s41586-020-2649-2" | python3 -m json.tool | head -40

Delete any reference that cannot be verified, and say what was removed.

Output template

# <Review title>
**Type:** narrative | scoping | rapid   **Coverage:** 2019-2025, English   **Searched:** 2025-10-14

## Question and scope
## Method (one paragraph): databases, strings in Appendix A, criteria, counts
   (identified 412 -> deduplicated 301 -> screened 301 -> full text 58 -> included 31)
## Findings by theme
### Theme 1 ... (claims with citations; strength words matched to evidence)
## Contested findings and why they differ
## Gaps and research opportunities (each tied to matrix evidence)
## Limitations of this review
## References (verified; DOI or stable URL for each)
## Appendix A: search log   Appendix B: synthesis matrix

Gotchas

• Seminal papers missing from the results almost always mean a vocabulary gap in the string, not that the paper is irrelevant. Test with gold papers.
• Review articles inflate apparent consensus: cite the primary studies for empirical claims and the review for the overview.
• arXiv-heavy fields: the same work appears as preprint, workshop, and conference paper. Count it once.
• Citation counts favor older papers and hype; never use them as a quality filter.
• “No studies found” is a finding only if the search was documented and tested.