# Session 2: Building the Human-AI Research Pipeline

URL: https://research-memex.org/docs/case-studies/systematic-reviews/session-2-ai-powered-practice
Description: A systematic review becomes easier to teach when the research pipeline is visible and testable.



*Building the pipeline once teaches more than reading about it ten times.*

This guide offers a hands-on workshop for designing and executing a systematic review with AI tools. The approach is cognitive-first: instead of just learning to operate tools, we learn to think with them. We engage with AI as a mirror that makes our own research processes visible, helping us deconstruct complex tasks like "finding a gap" or "building a theory" into explicit, repeatable steps.

By the end of this session, you'll move beyond generic AI interaction to purposeful orchestration.

## Learning outcomes [#learning-outcomes]

By the end of this session, you'll be able to:

* Construct a high-quality, curated literature set using discovery (Research Rabbit) and management (Zotero) tools
* Deconstruct complex research tasks (e.g., "synthesizing a framework") into explicit, step-by-step cognitive operations
* Design and implement structured, multi-step prompts (<Tooltip tip="Structured multi-step prompts that guide AI through complex research tasks">"[cognitive blueprints](/docs/implementation/core-references/cognitive-blueprint-prompts)"</Tooltip>) that guide AI through sophisticated analytical work
* Critically evaluate AI-generated outputs, identifying common failure modes (like "[botshit](/docs/implementation/core-references/failure-museum)" and paradigm blindness)
* Architect a research workflow that strategically combines your domain expertise with AI capabilities

## The complete research pipeline [#the-complete-research-pipeline]

```text title="diagram:research-pipeline"
     seed papers
          |
          v
  citation discovery
          |
          v
    expanded set
          |
          v
 +--------+--------+
 | Human curation  |
 +--------+--------+
          |
          v
  reference library
          |
          v
AI-assisted analysis
          |
 +--------+--------+
 | quality control |
 +--------+--------+
          |
  issues? +----> revise method
          |            |
          +<-----------+
          |
          v
 research synthesis
```

Human judgment provides quality gates throughout this pipeline: AI handles scale and pattern detection, while researchers provide curation and evaluation.

## Core readings: the AI toolkit [#core-readings-the-ai-toolkit]

### Mindset & mental models [#mindset--mental-models]

<Card title="Interpretive Orchestration" icon="book-open" href="https://doi.org/10.1177/14761270261448645">
  Why this matters: establishes the professional mindset for this approach. It frames the researcher as the person who designs and checks the workflow, not a passive operator.
</Card>

<Card title="&#x22;the void&#x22; by nostalgebraist" icon="brain-circuit" href="https://www.tumblr.com/nostalgebraist/785766737747574784/the-void">
  Why this matters: provides a mental model for context-setting. You are not simply "talking to an AI"; you are shaping a predictive system through the information you provide.
</Card>

### Techniques & best practices [#techniques--best-practices]

<Card title="The Prompt Report: A Systematic Survey of Prompt Engineering Techniques" icon="book-open" href="https://arxiv.org/abs/2406.06608">
  Why this matters: provides a shared vocabulary and technical map for prompt engineering practice.
</Card>

<Card title="Gemini 2.5 Pro Capable of Winning Gold at IMO 2025" icon="trophy" href="https://arxiv.org/abs/2507.15855">
  Why this matters: shows how strong results come from a strong process. It teaches [scaffolding](/docs/implementation/core-references/cognitive-blueprint-prompts): building a thinking process for the AI to follow.
</Card>

### Risks & responsibility [#risks--responsibility]

<Card title="Beware of Botshit: How to Manage the Epistemic Risks of Generative Chatbots" icon="triangle-exclamation" href="https://doi.org/10.1016/j.bushor.2024.03.001">
  Why this matters: gives researchers a framework for identifying epistemic risk in chatbot outputs.
</Card>

## Supplementary readings [#supplementary-readings]

### Quick reference guides [#quick-reference-guides]

<Card title="DAIR.AI — Prompt Engineering Guide" icon="info-circle" href="https://www.promptingguide.ai/">
  Why this matters: offers practical definitions and examples that complement the more formal "Prompt Report."
</Card>

<Card title="Anthropic — Prompt engineering best practices" icon="info-circle" href="https://docs.anthropic.com/en/prompt-library/library">
  Why this matters: moves from general theory to model-specific advice that you can test in your own workflow.
</Card>

### Foundational papers [#foundational-papers]

<Card title="Chain-of-Thought Prompting Elicits Reasoning in Large Language Models" icon="file-text" href="https://arxiv.org/abs/2201.11903">
  Why this matters: provides the research basis for breaking complex problems into reasoning steps.
</Card>

<Card title="ReAct: Synergizing Reasoning and Acting in LMs" icon="file-text" href="https://arxiv.org/abs/2210.03629">
  Why this matters: explains the Reason → Act → Observe loop behind many agentic systems.
</Card>

### Advanced context [#advanced-context]

<Card title="Andrej Karpathy — Context engineering thread" icon="lightbulb" href="https://x.com/karpathy/status/1937902205765607626">
  Why this matters: shifts attention from asking better questions to giving the model the information it needs to answer well.
</Card>

<Card title="Simon Willison — &#x22;In defense of prompt engineering&#x22;" icon="lightbulb" href="https://simonwillison.net/2023/Feb/21/in-defense-of-prompt-engineering/">
  Why this matters: explains why prompt engineering remains a useful research skill when it is treated as context design, not phrase tuning.
</Card>

## Session structure [#session-structure]

Our hands-on session will follow this structure:

1. Pipeline Overview - Understanding the complete workflow
2. Discovery & Curation - Mastering Research Rabbit and Zotero
3. AI Integration - Setting up Cherry Studio and MCP servers
4. Pipeline Practice - Working with a sample literature set

We use the scaling and scalability literature as a shared case study throughout these guides: research on how organizations grow, and on whether their structures, processes, business models, and capabilities can handle that growth. In this course shorthand:

> **Scaling**: growth in size or scope.

> **Scalability**: the capacity to grow without the organization breaking or losing coherence.

A seed library of 10 foundational papers is provided, which you can expand using the discovery methods we learn today.

### MCP server setup (Live demo) [#mcp-server-setup-live-demo]

During the AI Integration phase, we'll add the first MCP servers:

<Card title="Follow Along" icon="users">
  1. Open Cherry Studio → Settings → MCP Configuration
  2. Enable @cherry/filesystem to access your research files
  3. Add @cherry/sequentialthinking for structured analysis
  4. Test both servers with your sample papers
  5. Test what changes when AI has direct file access.
</Card>

For more MCP exploration, see the [MCP Explorer Guide](/docs/implementation/ai-environment-setup/mcp-explorer-guide).

By the end, you'll have a complete, tested workflow for systematic reviews conducted through human-AI collaboration.

## Pre-class setup for Session 2 [#pre-class-setup-for-session-2]

Before our hands-on session, please complete the following setup to ensure you're ready to dive in.

<Steps>
  <Step title="Complete initial setup">
    Follow the [Cherry Studio Setup Guide](/docs/implementation/agentic-ai-tools/cherry-studio-setup-guide) to complete Steps 1-6. This includes installation, API setup, and basic MCP configuration.
  </Step>

  <Step title="Test your environment">
    * Test at least one AI model to ensure it's responding.
    * Test the Zotero MCP integration to confirm it can access your library.
  </Step>

  <Step title="Prepare your research vault">
    * Set up your Obsidian vault with the recommended folder structure from the setup guide.
    * Bring 3-5 of your core "seed papers" as PDFs, ready to be added to your knowledge base.
  </Step>
</Steps>

During class, we will:

* Set up knowledge bases together
* Practice Zotero MCP searches
* Export conversations to Obsidian
* Create literature note templates
* Practice conversation forking for different analyses

## How Do You Know It's Working? [#how-do-you-know-its-working]

<AccordionGroup>
  <Accordion title="Success Indicators" icon="circle-check">
    * API responds within 5 seconds
    * Outputs include specific citations with page numbers
    * Tool connections verified with test queries
    * No error messages in console/logs
    * Workflow completes end-to-end without manual intervention
    * Output quality matches or exceeds manual work
  </Accordion>

  <Accordion title="Warning Signs" icon="triangle-exclamation">
    * Generic, vague responses ("very important," "significant impact")
    * Missing or incorrect citations
    * Suspiciously smooth narratives (no contradictions)
    * Error messages or connection timeouts
    * Repeated failures at same step
    * Output quality below manual baseline
  </Accordion>

  <Accordion title="Quick Diagnostic" icon="wrench">
    If something's not working:

    1. **Test connection:** Send simple query, verify response
    2. **Check logs:** Look for error messages
    3. **Verify credentials:** API keys valid and have credits
    4. **Restart tools:** Close and reopen application
    5. **Consult docs:** Search for specific error message
    6. **Ask for help:** Consult [llms.txt](/llms.txt) or email instructor
  </Accordion>
</AccordionGroup>

***

<Card title="Recommended Exercises" icon="lightbulb">
  1. Read the foundational papers on different review types (e.g., Llewellyn 2021, Yuki 2024).
  2. Develop a synthesis prompt: Using the provided sample papers, combine the IMO approach with a chosen paper's method.
  3. Prepare a presentation: Document notes for a 3-5 minute presentation on your synthesis.
  4. Continue expanding your personal literature library in Zotero using Research Rabbit.
  5. [MCP Explorer Challenge](/docs/implementation/ai-environment-setup/mcp-explorer-guide): Find 2-3 MCP servers relevant to your research on smithery.ai, install one, and test it.

  Note: the IMO paper provides a template for structuring AI thinking processes. Apply this to your own synthesis tasks.
</Card>

## Related Resources [#related-resources]

<IndexRows>
  <IndexRow title="Setup Guides" href="/docs/implementation/foundational-setup/quick-start-checklist">
    Complete tool installation and configuration
  </IndexRow>

  <IndexRow title="Core References" href="/docs/implementation/core-references/cognitive-blueprint-prompts">
    Prompt templates and model guides
  </IndexRow>

  <IndexRow title="Failure Museum" href="/docs/implementation/core-references/failure-museum">
    Learn from common AI limitations
  </IndexRow>

  <IndexRow title="AI-Readable Docs" href="/docs/introduction/llms">
    llms.txt and Markdown access for AI tools
  </IndexRow>
</IndexRows>

***

Navigation: [Return to Case Study Overview](/docs/case-studies/systematic-reviews/systematic-review-syllabus) • [Next: Session 3](/docs/case-studies/systematic-reviews/session-3-human-vs-ai-synthesis)