Scientific literature review

Scientific Literature Review

Build a reproducible evidence trail. Do not generate a polished narrative first and search for citations afterward.

Choose the review mode

Mode	Use when	Minimum output
Rapid evidence map	Time is limited; landscape and gaps matter	search log, source table, themes, gaps
Narrative review	Explain a field without claiming exhaustive coverage	scoped search, representative primary evidence, limitations
Scoping review	Map concepts, methods, or evidence types broadly	protocol, reproducible search, screening ledger, charted evidence
Systematic review	A focused question needs comprehensive, auditable selection	protocol, multi-database search, deduplication, dual/human screening plan, risk-of-bias plan, PRISMA records
Meta-analysis support	Comparable effect estimates exist	systematic-review foundation plus statistical protocol and human statistician review

Never call a review “systematic” merely because many sources were read.

Workflow

1. Frame the question with the appropriate structure: PICO/PECO for intervention or exposure, PCC for scoping, or a mechanistic question with explicit model, perturbation, comparator, and outcome.
2. Predefine dates, languages, study types, populations/models, outcomes, inclusion/exclusion criteria, and treatment of preprints. Record deviations.
3. Build concept blocks with synonyms, gene/protein aliases, controlled vocabulary, and spelling variants. Read references/database-search.md.
4. Search PubMed plus at least one complementary database appropriate to the topic. Preserve the exact query, database, filters, search date, result count, and PubMed Search Details/translation.
5. Export stable identifiers and deduplicate by DOI, PMID/PMCID, accession, then normalized title/author/year. Keep a duplicate-resolution log.
6. Screen title/abstract and then full text against the predefined criteria. For publication-grade or regulated work, a qualified human owns inclusion decisions and conflict resolution.
7. Extract study design, model/population, sample size, intervention/exposure, comparator, outcome, effect estimate, uncertainty, methods, funding/conflicts, and limitations into an evidence table.
8. Appraise evidence using a method appropriate to the study design; do not invent a universal score. Read references/evidence-appraisal.md.
9. Verify every citation against a primary bibliographic record before delivery. Read references/citation-verification.md.
10. Synthesize by evidence pattern, not paper-by-paper summary. Separate replicated findings, contested findings, methodological causes of disagreement, gaps, and inference.

Use the shared deep-research skill for parallel discovery when available, but retain this skill's scientific search log, source hierarchy, screening, and verification rules.

Non-negotiable quality rules

• Prefer primary papers for scientific claims; use reviews to orient and discover primary evidence.
• Label preprints, conference abstracts, retractions, corrections, and expressions of concern.
• Never infer full-text results from an abstract or search snippet.
• Quote sparingly and preserve the source's meaning; do not transform correlation into causation.
• Do not fabricate DOI, PMID, author, journal, year, sample size, effect, or quotation.
• Distinguish no evidence found from evidence of no effect.
• Report search and access limitations, unavailable full texts, and unsearched databases.
• Do not claim PRISMA compliance unless every applicable checklist item and flow record is complete; otherwise say “PRISMA-informed.”

Deliverable

Return:

1. question, review mode, scope, and protocol summary;
2. exact search log;
3. inclusion/exclusion and screening summary;
4. evidence table with stable identifiers;
5. synthesis with confidence and contradictory evidence;
6. limitations, gaps, and recommended next searches/experiments;
7. verified references;
8. reproducibility appendix with search date and tool/database versions when available.