Imported from diwad-code/getting_strange (
.agents/skills/serio-po-polsku-skill/.claude/skills/evidence-synthesis/SKILL.md). Install upstream withnpx skills add diwad-code/getting_strange --skill evidence-synthesis. Copyright stays with the author.
evidence-synthesis — integrate the evidence
Turn a screened, verified set of sources into a structured synthesis. Weight by
references/evidence-hierarchy.md; keep evidence and interpretation separate.
Choose the mode
- Narrative / thematic synthesis (default for reviews): organise by themes, mechanisms, or questions — not one-paragraph-per-paper. Show where studies agree, disagree, and why.
- Systematic review (PRISMA): when rigour and reproducibility are required.
Report the flow (identified → deduplicated → screened → eligible →
included) with counts, an evidence/characteristics table, and a
quality/risk-of-bias appraisal. Use the ready mermaid PRISMA diagram and
machine-readable search log in
references/reproducible-search.md.
Steps
- Group sources by theme or by PICO element.
- Tabulate key studies: design, sample, key finding, effect size, limitations, quality tier.
- Reconcile conflicts. When studies disagree, explain it via study quality, population, method, or date — don't just average. Higher-tier and replicated evidence dominates.
- State the bottom line per theme with calibrated certainty (see
scientific-consensus). - Surface gaps — what is unknown, underpowered, or unreplicated.
- Attribute everything. Every empirical statement carries a verified citation; your synthesis/inference is explicitly flagged as such.
Quality of the synthesis
- No cherry-picking: represent the body of evidence, including inconvenient results.
- Distinguish correlation from causation; don't over-generalise from narrow samples.
- Preprints and single studies are provisional; label them.
Output
A structured synthesis (themes or PRISMA sections) with an evidence table,
per-theme conclusions and certainty, identified gaps, and citations ready for
research-report.