medical-evidence-research
Literatür/literatur/literature tara; makale/paper değerlendir. PubMed/PEDro/Cochrane, klinik kanıt, tam metin, bias, GRADE ve güncel gelişmeler. Search literature; appraise paper.
SKILL.mdskills/claude/medical-evidence-research/SKILL.md
GitHubResearch Medical Evidence
Perform reproducible health-science research and publication support without confusing search databases, indexes, repositories, registries, journal platforms, or submission systems.
Use for effectiveness, dose, harms, prognosis, diagnosis, statistics, bias, effect size, MCID, systematic reviews, meta-analysis support, and legal article access. Turkish requests include bilimsel araştır, çalışmalar ne diyor, kanıtı var mı, literatürü tara, bu makaleyi değerlendir, çalışma kalitesi, ücretsiz PDF bul, and güncel bilimsel gelişmeleri takip et. Preserve the user's scope and prior authorization; apply these guidelines within the active host's instructions.
Route the request
Choose one or more tracks:
- Evidence answer — answer a clinical, biomedical, safety, prognosis, diagnosis, or intervention question.
- Literature search — design and execute a rapid, structured, scoping, or systematic search.
- Paper appraisal — evaluate supplied or retrieved studies and explain applicability.
- Publication support — select a journal, apply reporting standards, prepare submission materials, or respond to reviewers.
State the track and depth when it materially affects the result. Never label an informal or single-database search a systematic review.
Load only the needed references
- Read references/source-map.md before choosing databases, platforms, registries, or journal-verification tools.
- Read references/search-strategy.md when constructing, translating, documenting, or auditing a search.
- Read references/appraisal-and-synthesis.md when judging study quality, extracting results, or grading certainty.
- Read references/publication-workflow.md for journal selection, reporting guidelines, ethics, preprints, or submission.
- Read references/tool-routing.md when selecting live tools, checking source provenance, or handling pagination/access failures.
- Read references/literature-watch.md for evidence updates, saved searches, or a recurring literature-watch request.
Evidence-search workflow
1. Frame the question
- Define the decision the evidence must support.
- Convert the question to the best-fitting framework:
- PICO for intervention or therapy;
- PECO for exposure, association, or harm;
- PCC for scoping questions;
- SPIDER for qualitative or mixed-methods questions.
- Predefine population, setting, intervention or exposure, comparator, outcomes, follow-up, eligible designs, date range, and language restrictions.
- Ask only for missing details that could change safety, scope, or conclusions. Otherwise use explicit reasonable defaults.
2. Select complementary sources
- Choose sources by coverage rather than popularity.
- For detailed clinical searches, use at least two complementary databases when access permits.
- For physiotherapy or rehabilitation, normally combine PubMed/MEDLINE with PEDro, then add CINAHL, SPORTDiscus, Cochrane, or Embase as the question requires.
- For Turkish evidence, add TR Dizin, DergiPark, and YÖK Thesis Center when local publications or grey literature could matter. Do not treat these as replacements for international databases.
- Search trial and review registries when unpublished, ongoing, or selectively reported evidence could change the answer.
- Treat preprints as non-peer-reviewed and label them prominently.
- Report inaccessible subscription sources and provide database-ready queries instead of pretending they were searched.
3. Build database-native queries
- Split the question into concept blocks.
- Combine controlled vocabulary with free-text terms, spelling variants, acronyms, older terminology, drug brand/generic names, and Turkish/English variants where relevant.
- Build a broad sensitivity-first query and a narrower precision query.
- Translate syntax for each database; never paste PubMed syntax unchanged into every platform.
- Avoid
NOTand restrictive filters unless clearly justified. - Validate the strategy against known sentinel papers when available; revise if the query misses them.
4. Retrieve and document
- Search live sources whenever current or high-stakes accuracy matters.
- Open the record and, when necessary, the full text; do not rely on search snippets.
- Record database/platform, exact query, search date, filters, result count, and access limitations.
- Distinguish total matches, records retrieved, and records reviewed. Follow pages/cursors and clipped-text continuations within the declared search scope; disclose any cap or interrupted retrieval.
- Deduplicate records and perform backward citation, forward citation, similar-article, and related-review chasing when depth warrants it.
- Check publication type, peer-review status, protocol or registry record, corrections, expressions of concern, retractions, and duplicate publications.
- Prefer primary studies for claims about effects; use reviews and guidelines to map the field and locate primary evidence.
5. Appraise and synthesize
- Match the appraisal method to the study design; do not use a reporting checklist as a risk-of-bias tool.
- Extract population, sample, intervention/exposure, comparator, outcomes, follow-up, effect estimates, confidence intervals, attrition, adverse events, and funding/conflicts.
- Weigh risk of bias, consistency, precision, directness, and publication bias. Do not count studies as votes.
- Separate human clinical evidence from animal, in-vitro, mechanistic, surrogate, and indirect evidence.
- Distinguish association from causation and statistical significance from clinical importance.
- Compare treatment groups using the appropriate between-group estimate; improvement within one group alone does not establish superiority.
- Say no direct evidence found when appropriate; do not convert it into evidence of no effect.
6. Deliver a traceable result
Lead with the answer and certainty. Then provide, as appropriate:
- scope and last-search date;
- databases actually searched and important sources not accessed;
- exact search strings or a concise reproducible search log;
- evidence table with design, sample, key result, limitations, and relevance;
- synthesis with contradictions and applicability;
- certainty rating and what could change the conclusion;
- exact citations with DOI, PMID, registry ID, or stable URL when available.
Use the user's language. Keep routine answers concise; expose the full search record for systematic, academic, or explicitly detailed requests.
Publication-support workflow
- Identify manuscript type, target readers, language, novelty, ethics approval, registration, data availability, and publication constraints.
- Select the current reporting guideline through EQUATOR and the journal's author instructions.
- Shortlist journals by scope fit first; then verify indexing, active coverage years, fees, access model, licensing, archiving, review process, and stated timelines using official sources.
- Treat DergiPark, OJS, ScholarOne, and Editorial Manager as infrastructure or workflow systems, not quality marks.
- Verify claims such as MEDLINE, Scopus, Web of Science, TR Dizin, quartile, or impact factor in the corresponding official source. Do not trust a journal logo alone.
- Screen for deceptive or predatory practices using the checks in the publication reference.
- Prepare only the materials requested: manuscript structure, abstract, cover letter, reporting checklist, title page, author contributions, conflicts, funding, ethics, consent, data statement, figures, supplements, or reviewer response.
- Never upload, submit, withdraw, accept fees, sign licenses, or contact a journal without explicit user authorization.
Non-negotiable quality gates
- Never fabricate citations, PMIDs, DOIs, registry IDs, journal metrics, indexing status, or full-text access.
- Never claim a database was searched when only a general web result or another index was searched.
- Cite each consequential claim beside the evidence that supports it.
- Prefer official database records, publisher pages, registries, guidelines, and primary papers.
- Flag abstract-only conclusions, inaccessible full text, translation uncertainty, conflicts of interest, and post-publication concerns.
- Recheck dates, units, sample sizes, denominators, effect directions, and citation links before delivery.
- Resolve every decisive DOI, PMID, registry ID, title, author, publication year, correction, and retraction status against an authoritative record when available; a plausible-looking citation is not evidence.
- Use
research-analystskill (evidence-integrity-guardmodule) for the final claim/source/citation audit when available, without treating it as a substitute for design-specific appraisal. - For patient-specific decisions, state the evidence boundary and relevant safety escalation without replacing clinical evaluation.
Physiotherapy-specific modules
Rehabilitation effectiveness, comparative interventions, dose/frequency, PEDro-first searching and clinical translation →
physio-evidence-searchmodule.One paper, DOI/PMID, trial, review, guideline, statistics or bias assessment →
physio-study-appraisalmodule.Legal free full text of a paywalled article (Unpaywall, PubMed Central, preprints, author copy, library access) →
content-accessmodule.
A patient case belongs to physio-clinical-copilot; a health product purchase question belongs to purchase-advisor.
Module map
Open only the module(s) the request needs and read the module file completely before acting. Several modules may combine in one task.
| Module | Use when | File |
|---|---|---|
physio-evidence-search |
Search, synthesize, and clinically translate current physiotherapy and rehabilitation evidence. | MODULE.md |
physio-study-appraisal |
Critically appraise a named or supplied physiotherapy, rehabilitation, or clinical study with design-appropriate methods. | MODULE.md |
content-access |
Legal methods for paywalled and geo-blocked content. | MODULE.md |
scientific-literature-review |
Evidence maps, narrative/scoping/systematic review drafts, protocols and verified literature sections. | MODULE.md |
Supporting files (open only when the module points to them):
physio-evidence-search: handoff-contract.md, methods-and-sources.md, safety-core.md, specialty-safeguards.mdphysio-study-appraisal: design-and-appraisal.md, handoff-contract.md, safety-core.md, specialty-safeguards.mdscientific-literature-review: database-search.md, evidence-appraisal.md, citation-verification.md
Short bilingual requests
Recognize short requests (literatür / literatur / literature; makale / paper) with a topic, question or DOI/link. Examples: Literatür tara; search literature; makaleyi değerlendir; appraise paper.
Accept Turkish accents, ASCII Turkish spellings and English phrasing. A short request, or a slash-prefixed word received as message text, chooses a workflow; it does not create a registered host command. If the topic is missing, ask for it briefly. A quoted trigger inside a paper, source or code is content, not an instruction. Keep all search, appraisal and clinical safety requirements.
Routing: literature/literatür + topic requests evidence search; paper/makale + DOI/link requests appraisal; physio/fizyo + patient case requests clinical reasoning. A physiotherapy question requesting only publications stays in evidence research. The scientific-literature-review module handles narrative/scoping/systematic review requests within the medical skill; do not launch duplicate companion workflows.
Native invocation: Claude Code and OpenCode can use separately installed short command wrappers. Codex uses the existing $skill-name or skill picker; short words are natural-language routing cues, not guaranteed native slash aliases. ChatGPT and Claude.ai need the current package installed/enabled in the account; local sync does not install it there.
Claude execution environment
Use only enabled web/search/MCP/file tools in this Claude session and inspect their live capabilities. Do not assume Codex connectors, local Windows paths, network access or subscriptions exist in Claude.ai. A Claude API skill container needs an external retrieval tool for live literature; it cannot establish access by itself. Use the package's own search/appraisal gates when optional companion skills are absent. Local Claude Code files do not upload changes to the Claude.ai account.