How to use Claude for systematic literature reviews

Short recommendation
Use Claude to handle long-context PDF ingestion and structured extraction, but keep a reproducible pipeline: (1) prepare clean PDFs and metadata, (2) upload and chunk, (3) run structured extraction prompts that return JSON with explicit source pointers, (4) validate outputs against the PDF and export a traceable CSV/JSON. Save prompts, model version, and file hashes for auditability.

When to pick this approach (decision criteria)
- Best for: reviews with many full-text PDFs, complex qualitative synthesis, or when you need long-context reasoning. Claude’s longer context window helps keep document-level coherence.
- Avoid if: your budget is zero and you can’t use paid APIs, or you require strict regulatory/locked-down tooling (use institutional tools instead).
- Consider ChatGPT only if you need a simpler interface or your PDFs are few and short — otherwise Claude handles scale & context better.

Step-by-step Claude workflow (practical)
1) Preparation (local):
- Collect PDFs and metadata (title, authors, year, DOI, source URL). Use a reference manager (Zotero/EndNote) to keep IDs. Assign a unique doc_id to each file.
- Ensure searchable text (OCR where needed). Name files using doc_id (e.g., 0001_Smith2020.pdf).
- Create a manifest (CSV/JSON) with doc_id, filename, DOI, pages, and SHA256 file hash.

2) Ingest & chunk (Claude):
- Upload files to Claude (or the app wrapping Claude). If direct upload isn’t available, pre-chunk locally: 1–2 PDF pages per chunk or ~1,500–3,000 words. Smaller chunks reduce hallucination and keep provenance precise.
- For each chunk attach metadata: doc_id, filename, page_start, page_end, paragraph_index.

3) Structured extraction prompt (use a template):
- Ask Claude to return a machine-readable JSON array where each entry includes: doc_id, filename, page, paragraph_index, field(s) (title, design, population, n, outcome, numeric results with CIs/p-values if present), direct_quote (<=200 chars), and confidence_score (e.g., high/medium/low).
- Example fields: {"doc_id":"0001","page":4,"outcome":"HbA1c change","effect_size":"-0.5%","ci":"-0.8 to -0.2","quote":"…"}
- Force the model to only output JSON (no explanatory text) and to include the source pointer for every extracted claim.

4) QA & provenance checks
- Automatically compare returned page/quote against the original PDF text. Flag extractions where the quote does not exactly match.
- Keep a human reviewer sample (10–20% of records) and compute inter-rater agreement with the model’s extraction.

5) Synthesis & export
- Aggregate JSON into a CSV or database. Use the extracted numeric fields for meta-analysis; use the quotes and limitations fields for narrative synthesis.
- Save the manifest, prompts, model version/date, and output JSON for reproducibility.

Practical checklist (short)
- [ ] All PDFs OCRed and named with doc_id
- [ ] Manifest CSV/JSON created (DOI, URL, hash)
- [ ] Files uploaded or chunked with metadata
- [ ] Extraction prompt template implemented, JSON-only
- [ ] QA sampling plan and mismatch checks
- [ ] Exported dataset + saved prompts/model version

Best-for / Avoid-if
- Best-for: medium-to-large SRs, mixed methods synthesis, teams that need traceable quotes and page-level provenance.
- Avoid-if: single reviewer with no QA capacity, strict legal requirements unless you store all logs and hashes.

Notes on resources & team
- Budget affects whether you use a paid Claude tier or limit API calls. Prompt engineering and a small QA team (1–2 reviewers) drastically improve precision. For transparency, always store prompts, model version and file hashes.

If you want, I can give a ready-to-run JSON extraction prompt template and a sample CSV manifest structure to paste into Claude.

Compare Claude and ChatGPT