Run systematic literature reviews from a research question

Tools

Claude Code

Overview

Academic-physician-grade systematic literature review tool, built as a set of Claude Code skills.

Workflow

Academic-physician-grade systematic literature review tool, built as a set of Claude Code skills. Given a research question, it runs the full workflow of a professional systematic review — PICO formulation, multi-database search, PRISMA-grade dual-pass screening, citation chasing, quality appraisal, and structured synthesis — and produces an executive summary, evidence table, and Zotero-importable citations suitable for C-suite presentations and quality committee reports. It runs entirely inside your Claude Code subscription; no Anthropic API key is required (the Python backend is a pure data layer that talks to free literature APIs, never the model). Features PICO(S) formulation via a structured "Fundamental Four" interview, locked into a PROSPERO-style protocol before search begins Multi-source retrieval across PubMed, OpenAlex, and Semantic Scholar (Europe PMC optional), with CrossRef for DOI resolution and reference lists, and with a full MeSH + free-text query audit trail PRISMA 2020-compliant pipeline: deduplication, dual-pass abstract screening, backward/forward citation chasing, full-text eligibility assessment, and a generated PRISMA flow diagram Automated open-access PDF retrieval (Unpaywall, PubMed Central, Europe PMC, OpenAlex, CrossRef) with a guided manual step for the rest — no paywall bypass Study-type-aware quality appraisal: AMSTAR-2, Cochrane RoB 2, STROBE, COREQ, SANRA, or an Expert Opinion rubric, chosen automatically per article NHMRC evidence-level classification (I–IV) and structured gap analysis with an iterate-or-proceed decision point 6 human checkpoints so you can pause and resume across sessions without losing state Zotero-ready output: RIS + BibTeX citation files alongside the evidence table and narrative synthesis Optional polished export: bundle a completed review into a single PDF + DOCX deliverable Install Python 3.9+ (developed on 3.11; the test suite also passes on macOS system Python 3.9 — no CI enforces either). make deploy copies each skill in skills/ to ~/.claude/skills/{name}/SKILL.md, the folder structure Claude Code requires to discover a skill. Re-run it any time you pull skill changes. API Keys Setup config/api_keys.yaml (gitignored — never commit it) holds: Every key can also be set as an environment variable: NCBI_API_KEY, OPENALEX_EMAIL, CROSSREF_MAILTO, UNPAYWALL_EMAIL, SEMANTIC_SCHOLAR_API_KEY (the older S2_API_KEY name still works). None of these are required to run a review — they raise rate limits and add polite-pool access, and every source falls back to unauthenticated access. Semantic Scholar caveat: the shared unauthenticated pool answers HTTP 429 for long stretches. Without a key, retrieval retries with backoff and, if the pool stays saturated, reports Semantic Scholar as a degraded source in the retrieval summary (degraded_sources / source_status) and in search_log.md rather than recording a silent zero. A free key from https://www.semanticscholar.org/product/api removes the problem. Usage Open Claude Code in this project directory and invoke one of: The agent asks for your research question and guides you through every stage from there. If you close Claude Code mid-review, just start /lit-review again — it scans reviews/ for in-progress state and offers to resume. For a worked example of the full output — protocol, evidence table, executive summary, PRISMA flow diagram, gap analysis, and citation files — see reviews/demo-chf-care-transitions/. For guidance on which skill fits your situation, see docs/review_types.md. For full technical documentation, see docs/architecture.md. PDF Retrieval Full-text PDFs are needed for eligibility assessment (Stage 10) and data extraction (Stage 11). The agent gets you as far as it legitimately can, then hands the rest back to you: Automated, open-access only. After screening, the agent queries Unpaywall, PubMed Central, Europe PMC, OpenAlex, and CrossRef for a legally open copy of each included article and downloads what it finds. Manual, your own access. For articles with no open-access copy, the agent opens each one's page in your browser so you can download the PDF if you have legitimate access (personal subscription, institutional login already active in your browser, etc.), then imports what landed in your downloads folder. Library request. Anything still missing is written to library_request.csv — hand this to your own institution's library and request it through your normal channels.

Reported result

No measured outcomes or success metrics are reported in the source. The tool claims to generate PRISMA flow diagrams and structured synthesis, but actual performance on specific research questions is unknown. Proposed acceptance check: Verify that a test review generates a valid RIS file and a PRISMA diagram matching the input count, with zero silent failures in source retrieval logs.

Human role

The original post does not say how a person stayed involved.

Limitations

Prerequisites include Python 3.9+ and a Claude Code subscription. No Anthropic API key is needed, but free literature APIs have rate limits; Semantic Scholar may return HTTP 429 errors without a key. The tool only retrieves open-access PDFs automatically; users must manually download paywalled articles or submit library requests. Source reliability is secondary, as no execution evidence is provided.

Try it

I want to run a systematic literature review on [replaceable research question, e.g., 'effectiveness of telehealth for chronic pain'] using the lit-review-agent skills in Claude Code. First, install the skills via 'make deploy' and ensure Python 3.9+ is available. Start the /lit-review command. Guide me through the PICO formulation using the 'Fundamental Four' interview. Before executing the search, pause and show me the proposed PROSPERO-style protocol for my approval. Once approved, run the multi-source search across PubMed, OpenAlex, and Semantic Scholar. If Semantic Scholar returns 429 errors, log them as degraded sources but continue with other databases. After deduplication, present the abstract screening results for my dual-pass review. Do not proceed to full-text eligibility until I confirm the included studies. For any included study without an open-access PDF, open the browser link for me to download manually. Finally, generate the evidence table, PRISMA flow diagram, and RIS citation file. Save all outputs to [replaceable output directory]. Before finalizing, show me the executive summary and gap analysis for my review. If any step fails or inputs are missing, ask me for clarification rather than guessing.

More cases · Research & knowledge

• Draft a dissertation lit review from uploaded sources
• Run a 24/7 local agent that curates news digests
• Track daily arXiv papers into a Notion database
• Give a coding agent a paper corpus to find new techniques
• Run data cleaning to executive summary with agent pipeline