Week 5.5 - Building Your Research Workflow with Claude

What We'll Cover

So far this week, we've surveyed the landscape of AI literature tools and confronted the hallucinated citation problem. In this session, we shift from surveying tools to building something personal: a Claude-based research workflow tailored to your specific needs.

We'll cover three levels of customisation — from simple prompting techniques anyone can use today, through Claude Projects for ongoing research, to Claude Code's "skills" system for those who want maximum control. You don't need programming experience for any of this, but we will show you what's possible for those who want to go further.

The goal is practical: by the end of this session, you should have at least one new technique you can apply to your research today.

💬 Level 1: Better Prompting for Literature Review

The single biggest improvement most researchers can make is writing better prompts. A vague prompt produces a vague answer. A structured prompt — one that tells Claude exactly what you need, what you don't want, and how to handle uncertainty — produces dramatically better results. Below are four prompt templates you can start using immediately, whether you're on Claude's free tier or any other plan.

Note however, the first thing that you should do is to read the paper yourself. You won't always read the whole of every paper, but engage with it yourself first so that you can critically engage with the AI.

📑 The Structured Literature Prompt

Use this when you're starting research in a new area and need to understand the lay of the land.

PROMPT TEMPLATE I am researching [topic] for my [thesis/paper/review] in [discipline]. I need help understanding the current state of the literature. Please help me by: 1. Identifying the key themes and debates in this area 2. Suggesting search terms I should use (including terms I might not have considered) 3. Describing the major methodological approaches used in this field 4. Noting any recent shifts or emerging trends IMPORTANT: Do NOT generate specific paper citations. Instead, describe the landscape so I can search for papers myself using Semantic Scholar and Google Scholar. If you do mention specific researchers or papers, flag them clearly so I can verify they exist.

📊 The Synthesis Prompt

Use this after you've collected verified papers and uploaded them to Claude. This is where Claude shines — working with documents you've provided.

PROMPT TEMPLATE I have uploaded [N] papers on [topic]. Please: 1. Identify the key points of agreement across these papers 2. Identify the key points of disagreement or tension 3. Note any methodological differences that might explain different findings 4. Identify gaps — what questions do these papers not address? 5. For every claim you make, cite the specific paper(s) from my uploaded documents using [Author, Year] format

🔎 The Critical Reading Prompt

Use this when you need to deeply engage with a single paper, especially one that's central to your research.

PROMPT TEMPLATE Please read the attached paper and help me evaluate it critically: 1. What is the central argument or finding? 2. What methodology did the authors use, and what are its strengths and limitations? 3. What assumptions does the paper make (stated or unstated)? 4. How does this paper relate to [my specific research question]? 5. What questions would you ask the authors if you were a peer reviewer?

🛡️ Anti-Hallucination Prompt Strategies

These techniques reduce the risk of fabricated information. Use them in combination with any prompt above.

Explicitly forbid memory-based citations: "Do NOT generate citations from memory"
Demand uncertainty flagging: "If you are not sure about something, say 'I'm not certain' rather than guessing"
Scope to uploaded documents: "Based only on the documents I've provided..."
Normalise 'I don't know': "It's better to say you don't know than to fabricate an answer"
Separate discovery from analysis: Use Semantic Scholar or Elicit to find papers, then upload them to Claude to analyse them

Why does this work? LLMs respond to structure. When you give Claude a numbered list of tasks, it addresses each one. When you set explicit constraints ("do not generate citations"), it follows them far more reliably than if you simply hope it won't hallucinate. Think of prompting as giving instructions to a highly capable but very literal research assistant.

📁 Level 2: Claude Projects for Ongoing Research

Individual prompts are powerful, but they're ephemeral — each new conversation starts from scratch. If you're working on a research project over weeks or months, you need something persistent. That's where Claude Projects come in.

Claude Projects (available with a Claude Pro subscription) let you create a dedicated workspace for a research topic. You can upload your key papers, your methodology documents, your research proposal — and set custom instructions that persist across every conversation in that Project. Instead of re-explaining your context every time, Claude already knows who you are, what you're working on, and how you want it to behave.

Setting Up a Research Project

Here's how to structure an effective Claude Project for your research:

Create a Project with a clear name (e.g., "PhD — Water Governance in Cape Town")
Upload your key documents: your research proposal, 10–20 of your most important papers, your methodology chapter or plan, any relevant data descriptions
Set custom instructions that define how Claude should work with you (see the example below)
Start conversations within the Project — each one will have access to your documents and follow your custom instructions

The custom instructions are the key to making this work well. Here is an example you can adapt:

Example Project Instructions

CUSTOM INSTRUCTIONS You are my research assistant for my PhD on [topic] in [department] at [university]. My research question is: [question] My methodology is: [approach] When helping me: - Always ground your responses in the papers I've uploaded to this project - When I ask about the literature, cite specific papers from my project using [Author, Year] - Never generate citations from memory — if I need new papers, direct me to search Semantic Scholar or Google Scholar - When I ask you to summarise, always note what a paper's methodology was, not just its findings - Flag when my uploaded papers disagree with each other - Be direct about limitations and gaps - If I ask something that my uploaded papers don't address, say so clearly

Practical tip: Start small. Upload your 5 most important papers and your research proposal. As you find more relevant papers through your searches, add them to the Project over time. This creates a growing, curated knowledge base that Claude can draw on — and because every paper in it was found and verified by you, the citation problem is greatly reduced.

⚡ Level 3: Claude Code Skills (For the Adventurous)

This section is for those who want to go further. Most students won't need this right now, but understanding what's possible will help you decide how deep you want to go — and give you a sense of where these tools are heading.

What is Claude Code?

Claude Code is a terminal-based tool (or desktop app based) that runs Claude as an agent with access to your computer — it can read files, search the web, write code, and execute multi-step workflows. Unlike the web interface where you type one message at a time, Claude Code can take a complex request and break it into steps, executing each one and adapting as it goes. It's free to try with a Claude account.

You interact with it by typing commands in your terminal (the command line), and it responds by taking actions — reading your files, creating new ones, running searches, and reporting back what it found.

What is CLAUDE.md?

A CLAUDE.md file is a special Markdown file that acts as persistent instructions for Claude Code. It's read automatically at the start of every session — like giving Claude a "constitution" for how to work with you. You place it in your project folder, and every time you start Claude Code in that folder, it reads the file and follows those instructions.

Think of it as the equivalent of Project instructions (Level 2), but stored as a file on your computer rather than in Claude's web interface — which means you can version-control it, share it with collaborators, and customise it to a much greater degree.

Here is an example CLAUDE.md for a research project:

# Research Project: [Your Topic] ## My Research Context - PhD student in [Department], UCT - Research question: [Your question] - Methodology: [Your approach] - Stage: Literature review phase ## Rules for Literature Work - NEVER generate citations from memory - Always verify claims against uploaded sources - When I ask for papers, search Semantic Scholar API instead of guessing - Flag uncertainty with [UNVERIFIED] tags - Maintain my bibliography in Zotero-compatible format ## My Preferred Workflow 1. I provide a research question or theme 2. You help me formulate search queries 3. I run the searches in Semantic Scholar/Elicit 4. I upload the papers I find 5. You help me synthesise and critically evaluate them

What are Skills?

Skills are reusable, multi-step workflows you can trigger with a slash command (e.g., /lit-review). They live in a .claude/skills/ folder inside your project and can orchestrate complex research tasks that would otherwise require multiple separate prompts.

For example, a /lit-review skill could:

Ask you for your research question
Generate optimised search queries for Semantic Scholar
Help you evaluate the papers you find
Generate a structured synthesis with proper citations
Flag any claims that need verification

Instead of remembering and typing a long prompt each time, you simply type /lit-review and the skill guides you through the entire process.

The Teresa Torres Approach: Researcher Teresa Torres uses Claude Code with modular context files — separate files for different aspects of her research context, loaded as needed. Her daily workflow includes an automated research digest that surfaces new papers relevant to her predefined topics. This kind of setup takes time to build, but once in place it means new relevant research comes to you rather than you having to go looking for it.

🔒 A Note on Privacy and Data

⚠️ Important: When you upload papers to Claude (or any AI tool), those papers are processed on external servers. Before uploading anything, consider the following:

Check your university's data policy and your ethics approval before uploading sensitive or restricted data
For sensitive research data, consider using Claude's API with appropriate data handling agreements, or use on-premise solutions if available through your institution
NotebookLM (Google) and Claude Projects each have their own data policies — read them before uploading
Never upload papers that are under confidential review (e.g., manuscripts you are reviewing for a journal — these are shared with you in confidence)
Be cautious with unpublished data — if your research involves participant data, interview transcripts, or other sensitive material, check whether uploading it to a cloud service is compliant with your ethics approval

This is not a reason to avoid these tools entirely — it is a reason to be thoughtful about what you upload and where. Published papers are generally safe to upload. Your own writing drafts are fine. Sensitive participant data or confidential manuscripts require more care.

🎯 Choosing Your Level

There is no single right way to use AI in your research workflow. The best approach depends on how often you use these tools, what stage of research you're in, and how comfortable you are with technology. Here are the three levels summarised:

Everyone

💬 Better Prompts

Start today, no extra tools needed.

Works with Claude's free tier (or any LLM). Use the structured prompt templates from Section 1 to get dramatically better results from every conversation. The key principles: be specific, set constraints, separate discovery from analysis.

Best for: Everyone, regardless of experience level.

Regular Users

📁 Claude Projects

Requires a Claude Pro subscription.

Create a persistent workspace with your uploaded papers and custom instructions. Ideal for ongoing research where you want Claude to remember your context across multiple conversations. Build a curated knowledge base over time.

Best for: Researchers with an active, ongoing project who use Claude regularly.

Power Users

⚡ Claude Code + Skills

Free to try; requires comfort with the terminal.

Build automated, multi-step research workflows with CLAUDE.md and skills. Maximum flexibility and customisation. Your instructions live as files you control, and workflows can be shared with collaborators.

Best for: Researchers who want maximum control and are comfortable (or willing to learn) working in the terminal.

The most important thing is not which level you use — it's the principles behind all three: separate discovery from synthesis, ground everything in verified sources, build citation verification into your workflow from the start, and be explicit with AI about what you want and what you don't want it to do.

Summary & Key Takeaways

In this session, we moved from understanding AI tools to actually building with them. The three levels of customisation — better prompts, Claude Projects, and Claude Code skills — give you a pathway from simple improvements you can make right now to sophisticated workflows you can grow into over time.

Key takeaways:

Structure your prompts with numbered tasks, explicit constraints, and anti-hallucination safeguards — this alone will transform your results
Separate finding papers from analysing them: use Semantic Scholar, Elicit, or Google Scholar to discover papers, then upload verified PDFs to Claude for synthesis
Claude Projects give you persistent context — upload your key papers once and have Claude remember your research context across sessions
Claude Code and CLAUDE.md offer maximum customisation for those who want automated, repeatable research workflows
Always check data policies before uploading sensitive or confidential material to any AI tool
The principles matter more than the tools: verify sources, flag uncertainty, and never let AI-generated citations go unchecked

Next session: We put everything together with hands-on activities — a comparative search challenge, a citation verification exercise, and the weekly assessment. Come ready to work with your own research topic.