name: literature-review
description: >
Run a systematic biomedical literature review using the BioTeam-AI W1
pipeline. Searches PubMed and bioRxiv, screens and extracts data from
papers, synthesizes findings with anti-hallucination grounding, validates
citations against retrieved sources, and scores evidence via the RCMXT
framework. Supports MCP healthcare connectors for enhanced data access.
metadata:
author: bioteam-ai
version: "1.0"
domain: biomedical-research
allowed-tools: Bash(python:*) Read
Literature Review Skill
When to Activate
- User asks to "review literature" or "find papers" on a biomedical topic
- User asks about evidence for/against a hypothesis
- User wants to identify research gaps, contradictions, or novel findings
Prerequisites
- Python 3.11+ with the BioTeam-AI backend installed
ANTHROPIC_API_KEY and NCBI_EMAIL environment variables set
- Optional:
NCBI_API_KEY for higher PubMed rate limits
Usage
cd backend && uv run python -m app.skills.w1_cli --query "your research question" --budget 5.0
Pipeline Steps (13 total)
- SCOPE — Research Director defines search scope and strategy
- SEARCH — Knowledge Manager searches PubMed + Semantic Scholar (or MCP connectors)
- SCREEN — T02 Transcriptomics screens papers for relevance
- EXTRACT — T02 extracts structured data from included papers
- NEGATIVE_CHECK — Lab KB search for relevant negative results
- SYNTHESIZE — Research Director synthesizes findings (with anti-hallucination grounding)
- CONTRADICTION_CHECK — Ambiguity Engine detects contradictions
- CITATION_CHECK — Validates all citations against retrieved sources
- RCMXT_SCORE — Scores evidence claims (heuristic/LLM/hybrid)
- INTEGRITY_CHECK — Data integrity audit (gene names, retractions, stats)
- NOVELTY_CHECK — Knowledge Manager assesses novelty vs existing knowledge
- REPORT — Final output assembly with session manifest
Output Format
- JSON report with: executive summary, evidence table, contradiction matrix
- PRISMA flow data (identified → screened → included)
- RCMXT scores per claim (0.0-1.0)
- Citation validation report (verified/unverified counts)
- Session manifest with cost, model versions, step durations
Anti-Hallucination Guarantees
- SYNTHESIZE receives explicit paper list — only listed papers may be cited
- Citation validator cross-checks all citations against SEARCH results
- Unverified citations are flagged with warnings
- Empty-context queries labeled as "general knowledge"