Case Study: Systematic Reviews with AI

Systematic reviews teach orchestration because every step is inspectable.

Case study overview

This page outlines a course that teaches the Research Memex approach through a common academic task: the systematic literature review. This case study demonstrates one way to build AI partnership into the research workflow, from literature discovery through to final synthesis, while keeping every step inspectable.

We're sharing this approach to show how researchers can develop the complete pipeline, learning when and how to work with AI as a cognitive partner. Your implementation might look different based on your field and research questions.

Learning schedule & key topics

This case study is structured around four sessions, each building on the last:

Getting started with the case study

1. Download the Student Vault, which holds editable worksheets and a practice set.
2. Work through Session 1 in it, before any AI tool.
3. Follow the Quick Start Checklist as Sessions 2 to 4 add tools.

Each practical session has a detailed guide, and By Situation finds a page by what is in front of you when the sidebar alone does not. Adapt the pace and focus to match your own learning goals; instructors supply course-only readings separately.

Learning assessment

The learning process is assessed through two main components:

Learning through practice (50%)

• Session 1 benchmark: formative and not graded. Sessions 2 to 4 check their AI outputs against it.
• Session 2 Exercise: Master prompt development and cognitive scaffolding using a sample literature set.
• Session 3 Exercise: Develop critical evaluation skills by documenting AI failure modes and the limitations of automated synthesis.
• In-class work: Build presentation and peer feedback abilities.

Capstone learning project (50%)

Participants choose a final project that best serves their research goals:

• Option A: Validation skills - critically compare AI vs. human synthesis approaches
• Option B: Workflow design - develop reproducible human-AI research pipelines
• Option C: Quality control - build expertise in identifying and preventing common AI failure modes

Each option develops different competencies for research with AI partners. Choose what matters most for your work.

Tools & budget

• Essential tools: The workflows in this case study use Zotero (free), Research Rabbit (free), OpenAlex (free key), and Cherry Studio (open source) as the shared interface; The Research Workbench draws how they connect. Cherry Studio is the starting point because most participants arrive from a chatbot, and it lets them compare models side by side before an agent touches their files. Participants already comfortable with an agent can work in Claude Code or the Codex app instead; the desktop apps now cover most of what once needed a terminal, and the pipeline is the same.
• API budget: One paid account, several keys. Buy credits once on a gateway and issue each participant a named key with its own spend cap, so nobody handles payment and one dashboard shows who used what. Test the gateway before the course starts. Before spending credits, exhaust what is free: Google AI Studio's free tier, and whatever agent-capable model a subscription a participant already holds includes. The paid credits exist so that everyone can run the same task through different frontier models and feel the difference. Details in the API Keys Setup Guide.

Support

• Technical: See the setup guides for each tool.
• Content: See the PRISMA 2020 Guidelines and the Cochrane Handbook for systematic-review methodology.

What the first run taught us

The course became a case study because the first cohort taught its authors more than the syllabus did. Three things carried into the second run.

Setup is a session, not a prerequisite. The first tutorial had to become a setup workshop: Windows and macOS installs diverged, nobody had used Obsidian, and the written notes alone did not get anyone to a working agent. The most effective help came from the student who had just solved it, not from the instructor, who had forgotten what the first hour feels like. Plan for the workshop, and invite a graduate of the previous cohort to run it with you.

The agent team was built live. In Session 4 each participant brought the agent they had designed, and the class composed them into one team on the instructor's screen, one person at a time. That exercise was never in the written materials; it is now, in Session 4.

Most people went back to the chatbot. Six months after the course, most participants were using a chat window again, not even a single agent. That is the honest baseline, and the reason the second run aims to have participants run an agent team themselves rather than watch one. The tooling has moved in their favor: the desktop apps for Claude Code and Codex now cover most of what needed a terminal a year ago, and a host such as Kimi's Agent Swarm can take a research plan and design the team from it.

The second run, tutored by Raj Arasanal, added three things on the suggestion of the tutor and Prof. Autio: Session 1, a review done partly by hand before any AI so that later outputs have a benchmark; the placeholder convention on Review Prompt Templates, which marks what only the researcher can supply; and the one-prompt agent team in Session 4, which the tutor had run in class.

Navigation

Next steps:

• Session 1: A Review by Hand
• Session 2: Building the Human-AI Research Pipeline
• Session 3: Human vs AI Synthesis
• Session 4: Agentic Workflows

Resources:

• Quick Start Checklist

@misc{docs-case-studies-systematic-reviews-systematic-review-syllabus-2026,
  author = {Xule Lin},
  title = {Case Study: Systematic Reviews with AI},
  year = {2026},
  howpublished = {\url{https://research-memex.org/docs/case-studies/systematic-reviews/systematic-review-syllabus}},
  note = {ORCID: 0000-0001-7885-4194}
}