Use for topic overviews. Search across the main concepts and synthesize by theme, question, population, intervention/exposure, outcome, mechanism, or study design. Use search coverage and study tables only when they make the review clearer.
Comprehensive Evidence Map
Use when the user wants strong coverage. Track sources searched, query families, approximate records reviewed, key inclusion assumptions, and remaining gaps. Include study tables when many papers matter.
Formal Systematic Review Support
Use when requested or when the requested deliverable clearly requires publication-grade reproducibility. Define the review question, eligibility criteria, screening rules, extraction fields, risk-of-bias tools, and certainty framework; document the search reproducibly under Coverage Discipline.
Search Strategy
Search by concepts, not by the user's exact phrasing.
Extract the central concepts: condition, population, exposure/intervention, comparator, outcome, setting, mechanism, or study type.
Generate synonyms, abbreviations, spelling variants, older terms, disease subtypes, drug classes, and related mechanisms.
Combine synonyms with OR; combine concepts with AND.
Start broad enough to avoid missing the field, then narrow by design, population, outcome, date, or setting if results are noisy.
If results are sparse, remove the most restrictive concept, try synonyms, use broader disease families, and check references/reviews.
If results are dominated by irrelevant collisions, add disambiguating terms and exclusions carefully.
Use controlled vocabulary such as MeSH or Emtree only when the heading is known or verified.
Avoid filters that harm recall unless the user asked for them or the review type requires them.
For PubMed, prefer transparent queries that can be audited:
("concept one"[tiab] OR synonym[tiab])
AND
("concept two"[tiab] OR synonym[tiab])
Use MeSH only when appropriate:
("Heart Failure"[MeSH] OR "heart failure"[tiab] OR HFpEF[tiab])
AND
("Sodium-Glucose Transporter 2 Inhibitors"[MeSH] OR SGLT2[tiab] OR empagliflozin[tiab])
Do not repeat failed queries with cosmetic changes. Change the concept strategy.
Source Selection
Use any available source or tool that is academically or scientifically appropriate for the question. The list below is not exhaustive.
Prefer biomedical databases, trial registries, official agencies, guidelines, primary literature, established reviews, and scholarly discovery indexes for evidence claims.
PubMed/MEDLINE: default biomedical search.
Cochrane Library/CENTRAL: systematic reviews and controlled trials.
Embase: drugs, devices, pharmacovigilance, European indexing, conference abstracts when accessible.
ClinicalTrials.gov and WHO ICTRP: registered, ongoing, unpublished, or recently completed trials.
Professional society guidelines and official agencies: clinical-practice or public-health questions.
FDA, EMA, CDC, WHO, NICE, USPSTF, or equivalent bodies: regulatory, safety, surveillance, or guideline questions.
CINAHL, PsycINFO, Web of Science, Scopus, specialty databases: nursing, psychology, citation coverage, or field-specific questions.
Semantic Scholar, OpenAlex, Crossref, Google Scholar, citation chasing: broad scholarly discovery and forward/backward references; verify important claims against primary, review, guideline, registry, or official sources.
medRxiv/bioRxiv: preprints only; label as not peer reviewed.
If an important source is unavailable, state that limitation. Do not imply exhaustive coverage from one database.
Coverage Discipline
For rapid answers, briefly name what was searched.
For focused or comprehensive reviews, maintain enough search trace for audit:
For formal systematic review support, record exact search strings, platforms, dates searched, limits/filters, record counts, deduplication, screening decisions, and exclusion reasons.
Stop searching when:
the answer is adequately supported for the requested depth
additional query variants keep returning the same key studies
authoritative reviews/guidelines and citation chasing reveal no central missing sources
tool/time limits prevent more coverage and the limitation is explicit
the agreed formal-review search plan is complete
Never present "no evidence found" as proof of no effect. Say what was searched and what was not found.
Evidence Appraisal
Rank evidence according to the question. Treat these as judgment heuristics, not mechanical rules; directness, recency, quality, and fit to the question matter.
When foregrounding sources, prefer evidence that is direct, current, methodologically credible, and central to the field. Use lower-level evidence when it is the best available, and label why.
Intervention benefit: systematic reviews/meta-analyses, RCTs, pragmatic trials, then observational studies.
Harms and rare events: registries, pharmacovigilance, observational cohorts, case-control studies, case series when necessary.
Mechanism/basic science: model relevance, reproducibility, assay validity, dose/response, biological plausibility.
Guidelines: useful synthesis, not primary evidence; check the underlying evidence when the claim matters.
Separate these layers:
directly observed results
authors' interpretation
your synthesis
remaining uncertainty
Do not assign confident quality labels from abstract-only review. Say "abstract-level signal" or "full-text appraisal needed."
For the main bottom line, describe evidence strength in plain language: strong, moderate, limited, preliminary, or unclear. Explain the reason briefly. Do not treat this as formal GRADE unless formal certainty appraisal is requested.
Use formal tools only when the task calls for formal appraisal:
RoB 2 for randomized trials
ROBINS-I for non-randomized intervention studies
QUADAS-2 for diagnostic accuracy
AMSTAR 2 for systematic reviews
Newcastle-Ottawa Scale or another domain-accepted observational tool when appropriate
GRADE for certainty across a body of evidence
Screening And Study Tables
When screening candidate studies, use the user's criteria. If criteria are absent, infer provisional criteria from the question and label them provisional.
At title/abstract stage:
prefer maybe over exclude when uncertain
exclude only for clear mismatch or explicit exclusion trigger
give criterion-linked rationales
name what full-text evidence would resolve a maybe
When handling many records, deduplicate by stable identifiers first: DOI, PMID, trial registration ID, or another database ID. If identifiers are absent, compare exact title, then normalized title plus first author and year.
Use this table when screening matters:
| Study | Decision | Rationale | Confidence | Needs |
| --- | --- | --- | ---: | --- |
| Author Year | keep / maybe / exclude | criterion-linked reason | 0.0-1.0 | full text / outcome detail / population detail |