Systematic Literature Review Agent Skill

Systematic Literature Review

Overview

Produces a structured systematic literature review (SLR) across multiple

academic papers on a research topic. Given a topic query, searches arXiv,

extracts structured metadata from each paper, synthesizes themes, and emits a

final report with consistent citations.

Distinct from academic-paper-review: that skill does deep peer review of

a single paper. This skill does breadth-first synthesis across many papers.

> Poirot note: The original deer-flow skill uses a bundled

> scripts/arxiv_search.py + subagent task tool for parallel extraction.

> Poirot has neither, so this version uses bash with curl to the arXiv API

> directly + sequential single-agent extraction.

When to Use

• A literature survey on a topic ("survey transformer attention variants")
• A synthesis across multiple papers ("what do recent papers say about X")
• A systematic review with consistent citation format
• An annotated bibliography on a topic
• An overview of research trends in a field over a time window

Do not use when:

• User provides exactly one paper (use academic-paper-review)
• User asks a factual question (answer directly)

Workflow

Phase 1: Plan

Confirm with the user:

• Topic: the research area in plain English
• Scope: how many papers (default 20, hard upper bound 50), optional time

window, optional arXiv category (e.g. cs.CL)

• Citation format: APA, IEEE, or BibTeX (default APA)

If user says "50+ papers", cap at 50 and explain synthesis quality degrades

past that.

Phase 2: Search arXiv

Use bash with curl to the arXiv API. Extract 2-3 core keywords before

searching — don't pass the full topic description as the query.

# Search arXiv (use 2-3 core keywords, not the full topic)
curl -s "https://export.arxiv.org/api/query?search_query=all:transformer+attention&max_results=20&sortBy=relevance" | python3 -c "
import sys, xml.etree.ElementTree as ET, json
ns = {'a': 'http://www.w3.org/2005/Atom'}
root = ET.fromstring(sys.stdin.read())
papers = []
for entry in root.findall('a:entry', ns):
    papers.append({
        'id': entry.find('a:id', ns).text.split('/')[-1],
        'title': entry.find('a:title', ns).text.strip().replace('\n', ' '),
        'authors': [a.find('a:name', ns).text for a in entry.findall('a:author', ns)],
        'published': entry.find('a:published', ns).text[:10],
        'abstract': entry.find('a:summary', ns).text.strip(),
        'pdf_url': [l.get('href') for l in entry.findall('a:link', ns) if l.get('title') == 'pdf'],
        'abs_url': entry.find('a:id', ns).text,
    })
print(json.dumps(papers, indent=2, ensure_ascii=False))
"

Query tips:

• Use 2-3 core keywords, not the full topic phrase
• Use --category (arXiv cat: field) to narrow, not stuffing field names into query
• Always use sortBy=relevance (not submittedDate) for topical searches
• Run search exactly once; don't retry with modified queries

Phase 3: Extract metadata (sequential)

> Poirot note: The original skill delegates extraction to parallel

> subagents. Poirot has no subagents, so extract sequentially in your own

> context. For >20 papers, warn the user that sequential extraction is

> token-heavy and suggest splitting.

For each paper, extract from its abstract:

• arxiv_id
• title
• authors
• published_date
• research_question (1 sentence — what problem the paper tackles)
• methodology (1-2 sentences — how they tackle it)
• key_findings (3-5 bullet points)
• limitations (1-2 sentences)

Phase 4: Synthesize and format

Cross-paper synthesis — the report must do more than list papers:

• Themes: 3-6 recurring research directions across the set
• Convergences: findings multiple papers agree on
• Disagreements: where papers reach different conclusions
• Gaps: what the collective literature doesn't address

Citation formatting (inline, no bundled templates — format manually):

APA (default):

Author, A., & Author, B. (Year). Title. arXiv preprint arXiv:XXXX.XXXXX.

IEEE:

[1] A. Author and B. Author, "Title," arXiv preprint arXiv:XXXX.XXXXX, Year.

BibTeX (arXiv papers are @misc, not @article):

@misc{authorYear,
  title={Title},
  author={Author, A. and Author, B.},
  year={Year},
  eprint={XXXX.XXXXX},
  archivePrefix={arXiv}
}

Phase 5: Save and present

Save the full report to .poirot/outputs/slr-<topic-slug>-<YYYYMMDD>.md via

write_file. Present via present_files.

In the chat message, show a short preview:

• Executive summary — 3-5 sentence paragraph
• Themes list — bullet list of themes
• Paper count + file pointer

Do NOT dump the full report inline — per-paper annotations and references

belong in the file.

Report Structure

# Systematic Literature Review: [Topic]

## Executive Summary
[3-5 sentence overview]

## Methodology
[Search strategy, paper count, inclusion criteria]

## Themes
### Theme 1: [Name]
[Cross-paper analysis with citations]

### Theme 2: [Name]
[...]

## Convergences
[Findings multiple papers agree on]

## Disagreements
[Where papers diverge]

## Gaps
[What the literature doesn't address]

## Paper Annotations
### [Paper 1 Title]
- **Authors**: ...
- **Year**: ...
- **Research Question**: ...
- **Methodology**: ...
- **Key Findings**: ...
- **Limitations**: ...

### [Paper 2 Title]
[...]

## References
[Formatted per chosen citation style]

Pitfalls

• Query too specific: "diffusion models in computer vision" → 0 results.

Use 2-3 core keywords + category filter.

• sortBy=submittedDate: returns most recent papers in category regardless

of topic relevance. Use sortBy=relevance.

• Synthesis, not listing: A report that only lists papers one after another

is a failure mode. If you can't find themes, say so explicitly.

• >50 papers: synthesis quality degrades. Split by sub-topic.