| name | evidence-synthesis-forge |
| description | Orchestrates systematic reviews, scoping reviews, evidence maps, meta-analyses, umbrella reviews, and AI-assisted evidence synthesis. Use when designing protocols, eligibility criteria, search strategies, screening workflows, coding manuals, effect-size plans, synthesis reports, or reproducible evidence-review packages. |
Evidence Synthesis Forge
Use this skill as the general orchestrator for evidence synthesis. It helps move from a broad review idea to an auditable protocol, screening workflow, extraction plan, synthesis strategy, and report structure.
Core Principle
Separate:
- Question: what evidence is being synthesized.
- Search: how studies are found.
- Screening: how studies are included or excluded.
- Coding: how study features and effects are extracted.
- Synthesis: whether evidence is narratively summarized, mapped, or statistically pooled.
- Judgment: risk of bias, certainty, and interpretation.
Do not jump to meta-analysis before checking whether studies, outcomes, and effect sizes are comparable.
Intake
Identify:
- domain: environment, ecology, medicine, life science, economics, policy, education, psychology, or other;
- review type: systematic review, scoping review, evidence map, rapid review, meta-analysis, umbrella review, or second-order meta-analysis;
- question framework: PICO, PECO, PICOS, SPIDER, or custom;
- population/exposure/intervention/comparator/outcome;
- eligible study designs;
- desired outputs;
- whether machine learning will assist screening or extraction.
If the user is unsure, propose a review type and explain the tradeoff.
Workflow
- Build the review question.
- Define eligibility criteria.
- Draft search strategy and databases.
- Design screening stages and exclusion reasons.
- Define extraction fields and coding rules.
- Choose synthesis type: narrative, evidence map, first-order meta-analysis, umbrella review, or second-order meta-analysis.
- Specify risk-of-bias or quality assessment.
- Align the protocol and report with the relevant guidance source: PRISMA for reporting, Cochrane for intervention reviews, JBI for broader review types, CEE for environmental evidence.
- Add reproducibility artifacts: search log, screening log, coding sheet, analysis script, protocol.
Load:
references/review-types.md when the user needs help choosing review type or question framework.references/protocol-reporting-crosswalk.md when the user needs PRISMA/Cochrane/JBI/CEE alignment.templates/prisma-flow-counts.csv and scripts/generate_prisma_flow.py when the user needs a simple reproducible PRISMA-style flow diagram.
Routing
- For statistical pooling, use
meta-analysis-forge.
- For umbrella review or second-order meta-analysis, use
umbrella-review-skeptic.
- For machine-learning assisted screening or extraction, use
meta-ml-screener.
- For environmental, ecological, biomedical, or life-science reviews, use
environment-life-review-forge.
Output Modes
Protocol Skeleton
Use templates/evidence-protocol.md for full protocols.
Review Design Memo
Produce:
Review type:
Question framework:
Eligibility criteria:
Search plan:
Screening workflow:
Extraction fields:
Synthesis plan:
Bias/quality assessment:
Reproducibility artifacts:
Risks:
PRISMA Flow
Use templates/prisma-flow-counts.csv to record counts and scripts/generate_prisma_flow.py to generate a Mermaid diagram.
Count source:
Records identified:
Duplicates removed:
Records screened:
Reports assessed:
Studies included:
Studies in meta-analysis:
Known deviations:
Audit
When reviewing an existing protocol or review, focus on:
- vague eligibility criteria;
- incomplete search;
- unlogged exclusions;
- incompatible outcomes;
- missing risk-of-bias assessment;
- naive pooling;
- hidden ML decisions;
- overclaiming.
Guardrails
- Do not invent included studies.
- Do not invent search results.
- Do not invent effect sizes.
- Do not recommend statistical pooling when constructs or estimands are incompatible.
- Do not let machine learning replace final inclusion decisions without explicit protocol justification.
- Do not write conclusions before the protocol, search, screening, extraction, and appraisal logic are clear.
- Do not present PRISMA counts as final unless they are traceable to search, deduplication, and screening logs.
