Claude literature-review skill: 10 PRISMA pipelines (2026)

The prompt

Use the literature-review skill to scaffold a PRISMA 2020 review on 'transformer architectures for time-series forecasting'. Run the multi-database search (arXiv, Semantic Scholar, OpenAlex). Save the raw hits to ./review/01-identified.json, deduplicate to ./review/02-deduped.json, then screen titles+abstracts to ./review/03-screened.json. For each stage emit a count (n=) so I can render the PRISMA flow diagram with the scientific-schematics skill afterwards.

What slides.md looks like

# scripts/prisma_pipeline.py
import json, requests
QUERY = "transformer time series forecasting"
S2 = "https://api.semanticscholar.org/graph/v1/paper/search"

# 1. identification
r = requests.get(S2, params={"query": QUERY, "limit": 100,
    "fields": "title,abstract,year,authors,externalIds,citationCount"})
hits = r.json()["data"]
json.dump(hits, open("review/01-identified.json", "w"))
print(f"n_identified={len(hits)}")

# 2. dedupe by DOI then by title-hash
seen, uniq = set(), []
for p in hits:
    key = p.get("externalIds",{}).get("DOI") or p["title"].lower().strip()
    if key not in seen: seen.add(key); uniq.append(p)
json.dump(uniq, open("review/02-deduped.json", "w"))
print(f"n_deduped={len(uniq)}")

Pairs with

The prompt

Use the literature-review skill. Query Semantic Scholar's /graph/v1/paper/search endpoint for 'retrieval-augmented generation evaluation'. Sort the top 20 by citationCount descending. Return a markdown table with columns: Title | Year | Cites | Venue | DOI. Save raw JSON to ./review/s2-rag.json so I can re-rank locally if needed.

What slides.md looks like

import requests, json
URL = "https://api.semanticscholar.org/graph/v1/paper/search"
params = {"query": "retrieval augmented generation evaluation",
          "limit": 20,
          "fields": "title,year,citationCount,venue,externalIds"}
r = requests.get(URL, params=params, timeout=30); r.raise_for_status()
data = sorted(r.json()["data"],
              key=lambda p: p.get("citationCount", 0),
              reverse=True)
json.dump(data, open("review/s2-rag.json", "w"), indent=2)
for p in data:
    doi = p.get("externalIds",{}).get("DOI","–")
    print(f"| {p['title'][:60]} | {p['year']} | {p.get('citationCount',0)} | {p.get('venue','')[:24]} | {doi} |")

Pairs with

The prompt

Use the literature-review skill. For the query ("large language model" AND ("clinical" OR "medical")) AND year:2024-2026, hit arXiv (export.arxiv.org/api/query), OpenAlex (api.openalex.org/works) and PubMed (eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi). Merge by DOI, save to ./review/triple-merge.csv with columns: doi, title, source, year, citations.

What slides.md looks like

import requests, csv, urllib.parse as u
Q = '"large language model" AND ("clinical" OR "medical")'
endpoints = {
  "arxiv":    f"http://export.arxiv.org/api/query?search_query=all:{u.quote(Q)}&max_results=50",
  "openalex": f"https://api.openalex.org/works?search={u.quote(Q)}&filter=publication_year:2024-2026&per-page=50",
  "pubmed":   f"https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi?db=pubmed&term={u.quote(Q)}&retmode=json&retmax=50",
}
rows = []
for src, url in endpoints.items():
    rows += parse(src, requests.get(url, timeout=30))   # parse() returns dicts
# dedupe by DOI, fall back to lowercased title
seen, out = set(), []
for r in rows:
    k = r["doi"] or r["title"].lower();
    if k not in seen: seen.add(k); out.append(r)
csv.DictWriter(open("review/triple-merge.csv","w"), fieldnames=out[0].keys()).writerows(out)

Pairs with

The prompt

Use the literature-review skill. Take ./review/04-included.json, fetch BibTeX from doi.org for each entry (Accept: application/x-bibtex), normalize keys to AuthorYearKeyword format, write ./review/refs.bib. Emit a second file ./review/style-block.tex selecting the natbib style my journal wants (default: nature.bst).

What slides.md looks like

import json, requests, re
included = json.load(open("review/04-included.json"))
out = []
for p in included:
    doi = p["externalIds"]["DOI"]
    bib = requests.get(f"https://doi.org/{doi}",
        headers={"Accept": "application/x-bibtex"}, timeout=20).text
    # normalize cite-key: First-author surname + year + first-keyword
    first = p["authors"][0]["name"].split()[-1].lower()
    key  = f"{first}{p['year']}{p['title'].split()[0].lower()}"
    bib  = re.sub(r"@(\w+)\{[^,]+,", f"@\\1{{{key},", bib, count=1)
    out.append(bib)
open("review/refs.bib","w").write("\n\n".join(out))
open("review/style-block.tex","w").write(r"\bibliographystyle{nature}\bibliography{refs}")

Pairs with

The prompt

Use the literature-review skill on ./review/04-included.json. For each PDF, locate the Methods/Methodology section, extract: (a) sample size, (b) model architecture or technique, (c) dataset name, (d) primary evaluation metric. Write a markdown table to ./review/methods-table.md with one row per paper. If a paper has no PDF, fall back to its Semantic Scholar abstract.

What slides.md looks like

import json, requests
from pathlib import Path
included = json.load(open("review/04-included.json"))
rows = ["| Paper | n | Model | Dataset | Metric |", "|---|---|---|---|---|"]
for p in included:
    pdf = Path(f"pdfs/{p['paperId']}.pdf")
    text = read_section(pdf, heading="Methods") if pdf.exists() else p["abstract"]
    fields = claude_extract(text, schema={"n":"int","model":"str",
                                          "dataset":"str","metric":"str"})
    rows.append(f"| {p['title'][:40]} | {fields['n']} | {fields['model']} "
                f"| {fields['dataset']} | {fields['metric']} |")
open("review/methods-table.md","w").write("\n".join(rows))

Pairs with

The prompt

Use the literature-review skill on the top 5 papers in ./review/03-screened.json (sorted by citationCount). For each, extract its single-sentence contribution. Cluster the five contributions into themes. Then write a 200-word 'Gap analysis' section to ./review/gap.md naming three open questions that none of the five papers tackles — cite each claim.

What slides.md looks like

import json
top5 = sorted(json.load(open("review/03-screened.json")),
              key=lambda p: p.get("citationCount",0), reverse=True)[:5]

prompt = """For each paper, in 1 sentence: what is its central contribution?
Then group contributions into 2-3 themes.
Then list 3 open questions that NONE of the five address.
Cite every claim with [SmithYearKeyword]. Output markdown."""

# claude_call() is the literature-review skill helper
md = claude_call(prompt, attachments=[p["abstract"] for p in top5])
open("review/gap.md","w").write(md)

Pairs with

The prompt

Use the literature-review skill. Seed = DOI 10.48550/arXiv.1706.03762 (Attention is All You Need). Walk Semantic Scholar's /paper/{id}/citations endpoint two hops. Cap at 50 nodes per hop. Save the edge list to ./review/cite-graph.csv (src_id, dst_id, year) and a Graphviz DOT file to ./review/cite-graph.dot.

What slides.md looks like

import requests, csv
S2 = "https://api.semanticscholar.org/graph/v1/paper"
seed = "10.48550/arXiv.1706.03762"
edges, frontier, depth = [], [seed], 0
while depth < 2:
    next_front = []
    for pid in frontier:
        r = requests.get(f"{S2}/{pid}/citations",
            params={"limit": 50, "fields": "paperId,year"}).json()
        for c in r["data"]:
            cp = c["citingPaper"]
            edges.append((pid, cp["paperId"], cp.get("year","")))
            next_front.append(cp["paperId"])
    frontier, depth = next_front, depth + 1
csv.writer(open("review/cite-graph.csv","w")).writerows(
    [("src","dst","year"), *edges])
with open("review/cite-graph.dot","w") as f:
    f.write("digraph G {\n")
    for s,d,_ in edges: f.write(f'  "{s[:8]}" -> "{d[:8]}";\n')
    f.write("}\n")

Pairs with

The prompt

Use the literature-review skill. Seed paperId = 6b85b63579a916f705a8e10a49bd8d849d91b1fc. Hit /paper/{id}/references and /paper/{id}/citations. Save backward+forward into ./review/snowball.csv. Tag each row with origin='backward' or origin='forward' so I can audit which arm contributed which paper.

What slides.md looks like

import requests, csv
S2  = "https://api.semanticscholar.org/graph/v1/paper"
SEED = "6b85b63579a916f705a8e10a49bd8d849d91b1fc"

def fetch(arm):
    url = f"{S2}/{SEED}/{'references' if arm=='backward' else 'citations'}"
    r = requests.get(url, params={"limit": 200,
        "fields":"paperId,title,year,externalIds"}).json()
    key = "citedPaper" if arm=="backward" else "citingPaper"
    return [(arm, p[key]["paperId"], p[key]["title"],
             p[key].get("year",""),
             p[key].get("externalIds",{}).get("DOI","–"))
            for p in r["data"]]

rows = fetch("backward") + fetch("forward")
csv.writer(open("review/snowball.csv","w")).writerows(
    [("origin","paperId","title","year","doi"), *rows])
print(f"backward={sum(1 for r in rows if r[0]=='backward')} "
      f"forward={sum(1 for r in rows if r[0]=='forward')}")

Pairs with

The prompt

Use the literature-review skill on ./review/04-included.json. For each paper extract: first-author surname, year, method (one phrase), sample size or dataset, primary finding (one sentence). Output a sortable markdown table to ./review/summary-table.md AND a tidy CSV to ./review/summary-table.csv that I can pivot in pandas.

What slides.md looks like

import json, csv
included = json.load(open("review/04-included.json"))
header = ["author","year","method","sample","finding"]
rows = []
for p in included:
    fields = claude_extract(p["abstract"], schema={
        "method":"str","sample":"str","finding":"str"})
    rows.append({
      "author": p["authors"][0]["name"].split()[-1],
      "year":   p["year"],
      "method": fields["method"],
      "sample": fields["sample"],
      "finding": fields["finding"],
    })
# markdown
with open("review/summary-table.md","w") as f:
    f.write("| " + " | ".join(header) + " |\n")
    f.write("|" + "|".join(["---"]*len(header)) + "|\n")
    for r in rows: f.write("| " + " | ".join(str(r[h]) for h in header) + " |\n")
# csv
csv.DictWriter(open("review/summary-table.csv","w"),
               fieldnames=header).writerows(rows)

Pairs with

The prompt

Use the literature-review skill. Read ./review/{gap.md, summary-table.md, methods-table.md, refs.bib, prisma.png}. Write ./review/chapter-2.md with sections: 1) 250-word abstract, 2) Search strategy (Boolean string + databases + dates), 3) PRISMA flow (![](prisma.png)), 4) Summary table, 5) Methodology synthesis, 6) Gap analysis, 7) References. Use natbib \\citep keys from refs.bib.

What slides.md looks like

# scripts/stitch_chapter.py
import pathlib
chapter = []
chapter.append("# Chapter 2 — Systematic Literature Review\n")
chapter.append("## 2.1 Abstract\n" +
    pathlib.Path("review/abstract.md").read_text())
chapter.append("## 2.2 Search strategy\n" +
    pathlib.Path("review/search-strategy.md").read_text())
chapter.append("## 2.3 PRISMA flow\n![PRISMA](prisma.png)\n")
chapter.append("## 2.4 Summary of included studies\n" +
    pathlib.Path("review/summary-table.md").read_text())
chapter.append("## 2.5 Methodology synthesis\n" +
    pathlib.Path("review/methods-table.md").read_text())
chapter.append("## 2.6 Gap analysis\n" +
    pathlib.Path("review/gap.md").read_text())
chapter.append("## 2.7 References\n\\bibliography{refs}\n")
pathlib.Path("review/chapter-2.md").write_text("\n\n".join(chapter))
# pandoc -> PDF: pandoc chapter-2.md --citeproc --bibliography=refs.bib -o ch2.pdf

Pairs with