SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsliterature-review-generator

literature-review-generator

Writing & research

View repository

Updated 10-15 23:58

About this skill

Literature Review Generator

Overview

This skill is designed to produce literature reviews that are systematic, in-depth, and carry strong scientific argumentative weight. The agent will help map research development over time, identify contradictions/consensus, and logically build arguments that justify why the proposed new research is important to conduct.

Dependencies

  • literature-search-openalex
  • literature-search-arxiv
  • pubmed-database
  • literature-search-biorxiv
  • literature-search-europepmc

Quick Start

Example usage: "Use the literature-review-generator skill to map the development of reinforcement learning algorithms from 2018 to the present and find existing research gaps."

Workflow

1. Review Protocol Planning & Mode Selection

Determine the review boundaries (year range, document type) and choose one of the following working modes:

  • Narrative Literature Review Mode: For conceptual proposals or drafting general theory chapters.
  • Systematic Literature Review (SLR) Mode: Required for large-scale research (>100 papers). Uses formal inclusion-exclusion criteria, automated screening scripts, and a 4-stage PRISMA flow recorded into a checkpoint CSV (See the dedicated PRISMA Flow section below).
  • Thematic Review Mode: Groups papers by theme, concept, or problem sub-topic.
  • Chronological Review Mode: Traces the history of the development of ideas/technology from early pioneers to the latest trends to show paradigm evolution.
  • Critical Review Mode: Focuses on evaluating weaknesses in theoretical assumptions, evaluation weaknesses, data bias, and methodological limitations of previous studies.

2. PRISMA Flow & Checkpointing (Specifically for Large-Scale SLR Mode)

If the mode Systematic Literature Review (SLR) is chosen to handle searches in large quantities (thousands of papers), the agent MUST use programmatic screening methods (Programmatic Screening) to avoid manual LLM execution that wastes tokens. This 4-stage flow must be followed and saved in a CSV file (checkpoint):

  1. Identification (Fetch Metadata): Call free and open search engine APIs that are zero-configuration (literature-search-openalex, literature-search-arxiv, pubmed-database, etc.) to ONLY retrieve metadata (Title, Abstract, Authors, Year, DOI) from thousands of results in an integrated manner. DOWNLOADING FULL TEXT IS PROHIBITED. Save the raw output to prisma_1_identification.csv. GENERATING MOCKUP/DUMMY DATA IS STRICTLY PROHIBITED. You MUST use the requests module in Python to hit the real API endpoint or delegate the task to a real search agent. Apply a Retry/Backoff mechanism (maximum 3 retries if Timeoutoccurs). If it still fails, you must REPORT AN ERROR and STOP, not fabricate data.
  2. Automated Screening (Python Filter): Write and run a Python script (using Pandas) to screen prisma_1_identification.csv. Remove duplicates, limit the year range, and eliminate papers whose title/abstract text does not contain the main keywords at all. Save the remaining papers that pass the automated selection to prisma_2_screened.csv.
  3. Eligibility (Semantic & Quality Audit): For the remaining set of papers (50-100 papers), use an LLM to read the abstracts to ensure semantic relevance to Inclusion/Exclusion criteria. Use the agent source-quality-appraiser to discard low-ranking/predatory papers. Save the truly relevant papers to prisma_3_eligible.csv.
  4. Included (Full-Text Extraction): Download free PDFs/full texts (especially Open Access papers) from the final set (e.g. 20-40 papers) using the Core API or OpenAlex OA link, and use extract-methodology to dissect their parameters. Save the final synthesis matrix to prisma_4_included.csv.
  5. PRISMA Diagram: The agent is required to close this stage by generating a visual PRISMA flowchart in Mermaid diagramformat, which shows the reduction in the number of papers (n) at each stage (Identification -> Screening -> Eligibility -> Included).

3. Historical Search & Mapping (Timeline)

  • Search for seminal papers and the latest survey papers using literature search engines.
  • Arrange a chronological timeline of the development of that research.

4. Research Gap Analysis & Argument Map

  • Create an Argument Map to build a line of argumentation that justifies the position of the new research. Example flow:
    • Previous research has largely focused on the accuracy of methodology X.
    • However, such accuracy evaluations often do not test robustness under real-world conditions.
    • Several papers use imbalanced datasets without handling bias.
    • Therefore, this new research is present to close that gap by offering ablation-based robustness evaluation.

5. Literature Review Report Format

The structure of the literature review writing must include:

  • Introduction: The urgency of the topic, the research questions to be solved, and the review boundaries.
  • Literature Summary Matrix: A mandatory structured table with the following columns:
    Year Author Research Focus Method Dataset / Context Main Findings Main Limitations Relevance to Our Research
  • Argument Map (Research Position Map): A point-based explanation of the line of argumentation that leads to your research contribution.
  • Critical Analysis & Theme Categorization: Critical grouping of the literature based on the approach used (e.g., rule-based approaches vs deep learning).
  • Conclusion: Summary of theoretical and methodological implications for future research.

6. Citations & References

  • Ensure every claim is supported by appropriate references by including valid DOI/URL metadata in the Bibliography.

Common Mistakes & Critical Rules

  • Data Fabrication & DOI Forgery (FATAL): It is strictly forbidden to create fictitious papers merely to fulfill the PRISMA table. Creating a DataFrame containing fabricated paper titles is a fatal violation. Every paper must come from a real API and have a valid DOI link.
  • Forcing Users to Enter Complex API Keys: It is STRICTLY FORBIDDEN to display error messages that force users to configure an API Key manually if an API Key (such as Semantic Scholar) is not yet available. Instead, use free APIs that operate without a key (such as OpenAlex, PubMed, arXiv) or kindly direct users via chat to help them enter it if high quota is truly needed.
  • Summarizing Without Critiquing: Avoid the descriptive writing model of a list of summaries (e.g., "Researcher A found X. Researcher B found Y."). The agent must compare, group, and critique.
  • Random Chronology: Writing the history of a concept without a clear chronological order.
  • Invalid References: Writing citations without including a DOI or a traceable original reference link. Fabricating references is strictly forbidden.