Systematic Literature Review Pipeline

Overview

This skill runs a full systematic literature review from question to manuscript. It produces two deliverables: a PRISMA 2020-compliant .docx manuscript and a .xlsx review matrix with extracted data. The skill is optimized for social sciences research but adapts to any discipline.

The pipeline has seven phases. The user can enter at any phase depending on how far along they are — some users arrive with just a topic, others already have a stack of articles. The skill detects the entry point and picks up from there.

Phase Detection

Before starting, figure out where the user is:

If the user provides articles or references alongside their request, acknowledge them and integrate them into the appropriate phase rather than starting from scratch.

Phase 1: Research Question Formulation

The research question shapes everything that follows. A weak question produces an unfocused review.

PICO Framework (for intervention-focused reviews)

PCC Framework (for scoping or exploratory reviews)

Deliverable

A one-paragraph research question statement, formatted per framework:

PICO (systematic reviews):

Research Question: In [Population], how does [Intervention] compared to [Comparison] affect [Outcome]?

PCC (scoping reviews):

Research Question: What is known about [Concept] in [Context] among [Population]?

Plus 2-3 sub-questions if the review scope warrants them. If the user does not specify PICO or PCC, default to PCC for exploratory/scoping reviews and PICO for intervention-focused reviews.

Phase 2: Search Strategy Design

This phase produces a reproducible, documented search strategy. Read references/search-strategy-templates.md for database-specific syntax and Boolean patterns.

Step 2.1: Identify search terms

Break the research question into concept groups (typically one per PICO element). For each group, generate:

• Primary terms (from the research question)
• Synonyms and related terms
• Controlled vocabulary / subject headings (where applicable)
• Truncated forms (e.g., sustainab*, entrepren*)

Present as a concept table:

Step 2.2: Build search strings

Combine concept groups with AND:

("smallholder farmer*" OR "small-scale farmer*" OR "subsistence farmer*")
AND
("digital extension" OR "mobile advisory" OR "e-extension" OR "ICT-based extension")
AND
("crop yield" OR "agricultural productivity" OR "farm income" OR "technology adoption")

Step 2.3: Select databases

For social sciences, default to:

Adapt based on the user's discipline and institutional access.

Step 2.4: Define inclusion/exclusion criteria

Use 10 categories. Present as a table:

Deliverable

A Search Protocol Document containing: research question, concept table, full Boolean search strings for each database, inclusion/exclusion criteria table, and a search log template.

Handoff to Phase 3

After delivering the search protocol, explicitly tell the user to run the searches and return with results:

Your search protocol is ready. Next steps:

1. Run each search string in its target database (Scopus, Web of Science, etc.)
2. Export the results — BibTeX, RIS, or CSV format preferred; a list of titles and authors also works
3. Record how many results each database returned (for the PRISMA flow diagram)
4. Come back with the exported results and I'll start screening

If you need help accessing databases, check your university library portal.

Do NOT proceed to Phase 3 without user-provided search results. The skill cannot run database searches itself.

Phase 3: Study Selection & Screening

Step 3.1: Establish the starting pool

Work with what the user provides:

• If user provides exported references (BibTeX, RIS, CSV): parse and use as the starting pool
• If user provides a list of articles (titles, authors, years): use those
• If user provides PDFs or summaries: extract metadata and use those
• If user provides nothing: return to Phase 2 and deliver the search protocol with the handoff instructions above

Step 3.2: Two-stage screening

Stage 1 — Title & Abstract:

• Apply inclusion/exclusion criteria
• When uncertain, include for full-text review
• Document reason for each exclusion

Stage 2 — Full-text:

• Apply detailed eligibility criteria
• Document specific exclusion reasons (these go in the PRISMA flow diagram)

Step 3.3: Document PRISMA flow numbers

Track and record:

• Records identified from databases (per database)
• Records from other sources
• Duplicates removed
• Records screened (title/abstract)
• Records excluded at screening
• Full-text articles assessed
• Full-text articles excluded (with reasons)
• Studies included in final synthesis

Deliverable

A populated PRISMA flow diagram (as a formatted table in the manuscript) and a screening log.

Phase 4: Data Extraction

Step 4.1: Design extraction form

Create a review matrix with these standard columns (adapt based on the review topic):

Step 4.2: Populate the matrix

Extract data from each included study into the matrix. If the user provides article content (PDFs, summaries, notes), extract directly. If not, work from what's available (titles, abstracts, user descriptions).

Deliverable

A .xlsx review matrix file. See "Review Matrix (.xlsx) Generation" below for technical details.

Phase 5: Quality Assessment

Select tool based on study designs

Assessment process

For each included study:

1. Identify the study design
2. Apply the appropriate quality assessment tool
3. Rate each domain (Low risk / Some concerns / High risk, or equivalent)
4. Assign overall quality rating
5. Add the score to the review matrix

Deliverable

Quality assessment results added to the review matrix, plus a summary table showing risk of bias across studies.

Phase 6: Data Synthesis

Using Quality Scores from Phase 5

The quality assessment results from Phase 5 must inform the synthesis — do not ignore them:

• High-quality studies (Low risk of bias): treat as primary evidence. Lead with these findings.
• Medium-quality studies (Some concerns): include but note the limitation. Use phrases like "though limited by [specific concern], this study found..."
• Low-quality studies (High risk of bias): do NOT exclude silently — report their findings but weight them lower. Flag the quality concern explicitly: "However, this finding should be interpreted with caution due to [specific quality issue]."
• If all included studies are low quality, state this as a review-level limitation in the Discussion section.

Narrative synthesis (default for social sciences)

Organize findings thematically rather than study-by-study. The synthesis should:

1. Identify themes that emerge across studies
2. Compare and contrast findings within each theme
3. Note patterns — where do studies agree? Where do they diverge?
4. Explain heterogeneity — why might studies find different things? (Quality differences from Phase 5 may explain divergent findings)
5. Map gaps — what hasn't been studied?

Synthesis structure pattern

For each theme:

• State the theme and how many studies address it
• Present the consensus findings (with citations)
• Present the divergent findings (with citations)
• Discuss possible explanations for differences
• Connect to the research question

Quantitative synthesis (if applicable)

If studies are sufficiently homogeneous and report comparable effect measures, consider:

• Pooled effect estimates
• Forest plots
• Heterogeneity statistics (I², Cochran's Q)
• Sensitivity analyses

Most social sciences reviews use narrative synthesis. Only attempt meta-analysis if the user's studies are appropriate for it.

Deliverable

A structured synthesis narrative organized by themes, ready for integration into the manuscript.

Phase 7: Manuscript Production

This phase produces the final .docx manuscript.

Document Generation Strategy

There are two pathways to produce the .docx. Use Pandoc (primary) when the user has Pandoc installed and wants automatic citation rendering. Use docx-js (fallback) when Pandoc is not available or when fine-grained formatting control is needed.

If using Pandoc, the docx-js formatting specs below serve as reference values for a Pandoc reference template (if the user wants to match the same look).

Manuscript structure (PRISMA 2020)

Writing standards

Academic voice:

• Third person unless discipline conventions differ
• Hedging for uncertain claims ("suggests," "indicates," "appears to")
• Active voice for clarity where appropriate
• Define technical terms on first use

Citation density:

• Introduction: moderate
• Methods: low (citing methodology sources)
• Results: high (every finding attributed to source studies)
• Discussion: moderate (comparing with external literature)

Citation format — APA 7: Read ../../references/apa7-guide.md for the complete reference.

Review Matrix (.xlsx) Generation

Use openpyxl to create the review matrix spreadsheet.

Structure

Sheet 1: "Review Matrix" — the main data extraction table with all columns from Phase 4.

Sheet 2: "Quality Assessment" — risk of bias / quality scores per study per domain.

Sheet 3: "PRISMA Flow" — the flow diagram numbers in tabular form.

Formatting

from openpyxl import Workbook
from openpyxl.styles import Font, PatternFill, Alignment, Border, Side

# Header row: bold, dark background, white text
header_fill = PatternFill('solid', fgColor='1B4332')
header_font = Font(name='Arial', size=11, bold=True, color='FFFFFF')

# Data rows: alternating light fills for readability
alt_fill = PatternFill('solid', fgColor='F0F7F4')

# Column widths: set generously for readability
# ID: 8, Authors: 25, Year: 10, Title: 40, etc.

After creating, recalculate formulas if any are used:

python scripts/recalc.py review_matrix.xlsx

Manuscript Formatting Reference (docx-js fallback)

These specifications are used when generating .docx directly via docx-js (Node.js, npm install -g docx). When using Pandoc, these serve as reference values for matching the layout.

Page setup

// Portrait, US Letter, 1-inch margins
{ width: 12240, height: 15840, margin: { top: 1440, right: 1440, bottom: 1440, left: 1440 } }

Typography

• Body: Arial, 12pt (size: 24 half-points)
• Line spacing: double (line: 480)
• Paragraph spacing: 200 after
• Headings: Bold, hierarchical (H1: 16pt, H2: 13pt, H3: 12pt)

Reference list

• Hanging indent: indent: { left: 720, hanging: 720 }
• Alphabetical by first author surname
• Include DOI as hyperlink where available
• Font size: 11pt (size: 22)

PRISMA flow diagram (table format)

Render the PRISMA flow as a structured table in the manuscript with four sections (Identification, Screening, Eligibility, Included), using cell shading and borders to create a clean visual flow.

Tables in the manuscript

Include at minimum:

1. Table 1: Characteristics of included studies (summary of review matrix)
2. Table 2: Quality assessment summary
3. PRISMA Flow Diagram (as formatted table)

Zotero Integration (Citation Helper)

Use the /scholar:citation-helper skill to search the researcher's Zotero library (the Zotero CSL-JSON library configured in the plugin's .local.md file (default: references/salsu-library.json)) throughout the review process.

Integration Points by Phase

Citation Workflow During Manuscript Writing

1. Insert Pandoc citation keys: Use [@citekey] from the CSL-JSON id field
2. Multiple sources: [@key1; @key2; @key3]
3. Narrative citations: @citekey (e.g., @moore.westley2011 found that...)
4. Generate reference list: Use /scholar:citation-helper Operation 4 after writing is complete

Source Lookup Priority

0. ZOTERO: Search the Zotero CSL-JSON library configured in the plugin's .local.md file (default: references/salsu-library.json) by author, title, keyword, year
   ├── Match found → Use [@citekey] — researcher already has this source
   └── No match ↓
1. LOCAL PDFs: Search the local PDFs directory configured in .local.md for relevant papers
   └── No match ↓
2. WEB SEARCH: Search for additional academic sources (fallback)

Pandoc Rendering

To produce a formatted .docx with resolved APA 7 citations and automatic reference list:

# Replace <path-to-csl-json-library> and <path-to-apa-csl> with your configured paths from .local.md
pandoc manuscript.md \
  --citeproc \
  --bibliography=<path-to-csl-json-library> \
  --csl=<path-to-apa-csl> \
  -o manuscript.docx

Process Summary

Phase 1: Research Question  →  PICO/PCC framework
         ↓
Phase 2: Search Strategy    →  Concept table, Boolean strings, database selection
         ↓                     + search Zotero library for existing sources
Phase 3: Screening          →  Title/abstract + full-text, PRISMA flow numbers
         ↓
Phase 4: Data Extraction    →  Review matrix (.xlsx), use Zotero metadata
         ↓
Phase 5: Quality Assessment →  Risk of bias per study, added to matrix
         ↓
Phase 6: Synthesis          →  Thematic narrative synthesis with [@citekey] citations
         ↓
Phase 7: Manuscript         →  PRISMA-compliant .docx (Pandoc --citeproc for citations)

Reference Files

• ../../references/apa7-guide.md — Complete APA 7 formatting rules for citations and reference lists
• ../../references/prisma-checklist.md — PRISMA 2020 reporting items checklist with flow diagram template
• references/search-strategy-templates.md — Database-specific search syntax (Scopus, WoS, PubMed, ERIC) and Boolean patterns
• ../../references/review-methodology.md — Detailed methodology guide covering all seven phases with examples and common pitfalls