Analyze Qualitative Data with AI
A free, self-paced mini-course on using AI rigorously to analyze user interviews, customer feedback, and open text responses. Learn the prompting strategies that enforce solid methodology and produce results you can trust.
- Investment
- Free
- Format
- Self-paced · 9 lessons · ~1 hour
- Enrollment
- Open enrollment
Instant access. Learn at your own pace.
Lifetime access to lessons and materials
Revisit every lesson, the prompt, and any resources at any time.
The complete structured prompt, ready to use
Get the full five-step prompt and an XML-tagged version to use on your own projects right away.
Certificate of completion
A shareable credential when you finish.
Conduct AI-supported analysis you can actually trust.
It's not a magic prompt. It's not a special tool. The secret to doing rigorous qualitative analysis with AI is even more simple: prescribe multiple, discrete analytic steps.
This mini-course teaches the skills and techniques that enable any AI tool to conduct high-quality, reliable analysis of qualitative data.
The course is built around one structured prompt, and you'll learn it the way you'd learn any solid analytical method: not just the steps, but why each one exists. That means understanding how to sequence the AI's work so each stage builds on the last instead of racing to a conclusion, how to keep a model's judgment calibrated to yours before it ever touches your real data, and how to build in the kind of traceability that lets you stand behind every pattern you report (quote, participant, and all).
You'll also learn to read the output like a researcher, not just a consumer. That means recognizing the specific ways AI-assisted analysis tends to go wrong (categories that multiply past usefulness, interpretations that outrun the evidence, quotes that quietly lose their source) and knowing the exact move that corrects each one. By the time you're done, evaluating AI output critically will feel as natural as evaluating a research assistant's first pass at coding.
-
Build a structured, step-by-step AI prompt that enforces sound analytical practice instead of letting the AI race to conclusions in a single pass.
-
Manage context window limits so analysis doesn't silently drift or degrade as your dataset grows.
-
Apply role assignment, traceability requirements, and an alignment step so the AI's output is grounded in your data and matched to how you're thinking about the project.
-
Evaluate AI analysis output for the specific error patterns: category sprawl, overinterpretation, lost traceability.
-
Customize the prompt for your own data type and question, adding steps without breaking the checkpoints and traceability built into the original.
Good fit if…
- You've pasted interview transcripts into an AI tool and gotten output that looked good but felt off, and you want to know why and what to do instead.
- You analyze interviews, customer feedback, open-ended survey responses, or support tickets, and you want a structured, repeatable AI workflow, not just a prompt to copy and hope works.
- You're responsible for making sense of customer interviews, open-ended survey responses, support tickets, or sales call recordings, and you need to be able to stand behind the patterns you surface.
Not a fit if…
- You want AI to do the work while you step away from the data. This workflow keeps the researcher actively involved at every checkpoint.
- You're new to working with qualitative data and looking for a methods foundation. This assumes you already have some experience coding or categorizing data.
- You need to build a production AI system or pipeline. This is about analysis workflow for researchers, not engineering.
You'll need
- You've worked with messy qual data before: interviews, open-ended survey responses, usability sessions, support tickets, or other unstructured customer language.
- You have access to an AI chat tool like Claude, ChatGPT, or any other LLM chat interface.
- You're interested in building your own workflow rather than relying on pre-fab ones in research tools.
Helpful, not required
- A real dataset of transcripts or open-ended responses you want to practice with. (The course also provides sample data from a fictitious autonomous vehicle rideshare project.)
- Some familiarity with codebook-based or thematic analysis, even if you've mostly done it informally.
- An upcoming project where you expect to need to analyze a lot of qualitative data.
9 lessons. One structured prompt.
Cognitive Scaffolding
Apply cognitive scaffolding so each step's output becomes the context for the next, rather than asking AI to make every analytical decision at once.
You'll be able to
- Explain why a sequenced prompt produces dramatically better analytical output than a single "analyze this" request
- Describe how accumulated context at each step constrains and improves what the AI produces next
- Connect scaffolding to how a human researcher actually works: read, build an initial codebook, test it, refine it, apply it, then interpret
The Context Window
Manage context window limits so analysis doesn't degrade as your dataset grows.
You'll be able to
- Identify the context window of the AI tool you're using and verify whether your dataset fits
- Recognize the three ways overflow shows up: the prompt falls out first, then the beginning of your data, then the analysis drifts from both
- Apply the three tactics: batch your data, re-paste key instructions between messages, and start a new chat with the finalized codebook when needed
Role Assignment
Choose a role assignment that activates the right analytical lens for your specific project, not just a generic title.
You'll be able to
- Explain why "You are a Senior UX Researcher" and "You are a qualitative researcher" produce meaningfully different output from identical data
- Apply the rule: if changing a word or phrase would lead to a very different result, invest care in that word
- Select a role matched to your analytical question (usability, revenue insights, behavioral economics) rather than defaulting to your own job title
Traceability
Build exact-quote traceability into every claim so any pattern can be traced back to a specific participant and timestamp.
You'll be able to
- Write traceability rules into the prompt that require exact quotes, participant IDs, and timestamps, never paraphrases presented as quotes
- Apply attribution order: ask for the source location before the quote to increase accuracy and reduce fabrication
- Use the [NEEDS HUMAN REVIEW] flag to surface genuinely ambiguous passages rather than pushing AI to categorize them on its own
Alignment
Run an alignment calibration step with seed examples so the AI's analytical approach matches yours before any data is analyzed.
You'll be able to
- Provide 3–5 diverse seed examples with rationale, covering the full range of what you're listening for in the data, not similar examples that teach the AI a narrow lens
- Ask the AI to reflect back why you coded each example, not just to copy your codes, so you can catch misalignment before it compounds
- Choose code names deliberately: the AI weights these heavily, and superficial language pulls the output in a different direction than you intend
The Prompt: How to Set Up and Run It
Set up and run the full five-step prompt on your own data, working the stop points as intended rather than skipping past them.
You'll be able to
- Prepare the four inputs before pasting the prompt: project context, seed examples, your AI tool's context window size, and your data
- Work each STOP point as a genuine checkpoint: read the output, push back where the rationale doesn't match your sense of the data, and correct before the next step
- Redirect the AI if it tries to skip or merge steps: "Go back to Step X and complete it before moving on"
What Good Output Looks Like (and What to Watch For)
Evaluate analysis output for the quality signals that show the process is working and the four problems that most researchers catch too late.
You'll be able to
- Identify the four quality signals: distinct codes, rationale that matches the data, varied confidence scores, and [NEEDS HUMAN REVIEW] flags that show the AI is entertaining ambiguity
- Catch the four problem signals: category sprawl, favorite codes, overinterpretation, and lost traceability, and apply the specific correction for each
- Add a separate verification step when traceability breaks down, or use a second model as an independent checker
How to Customize this Prompt for Your Work
Extend the base prompt with new analytical steps and interpretive checkpoints suited to your own workflow and project types.
You'll be able to
- Add theme generation and negative case analysis as new steps, kept separate from coding so what's analytically true doesn't get conflated with what the business wants to hear
- Build in transformational moves: reflective moments, stakeholder feedback prompts, member-checking steps, and researcher debriefs
- Use contextual documentation to simulate stakeholder perspectives on your output, as a complement to, not a substitute for, looping in actual stakeholders
Using XML Tags to Improve AI Performance Bonus
Use XML markup to make complex, multi-step prompts more reliable, especially with longer datasets and nested instructions.
You'll be able to
- Understand what XML tags do structurally: they make the boundaries between instructions, data, and output formats explicit rather than relying on whitespace the AI may misread
- Apply the three-level hierarchy used in the XML version of the prompt: workflow, step, and step components (goal, action, exit rule)
- Decide when the XML version is worth the added editing complexity versus using the plain-text version, and how to preserve the tag hierarchy when customizing
Myth
AI isn't great with nuance.
Busted
LLMs turn each token into a vector with 4,096 dimensions, each of which represents some different dimension of a word's meaning. Can you do that?
Rephrase
LLMs are highly linguistically sensitive, but they don't perceive nuance the same way humans do.
Myth
There's a single magic prompt trick that will get me better results.
Busted
An LLM is one of the most complex machines ever built. Pulling one lever is not going to change anything dramatically.
Rephrase
Building in multiple prompt engineering techniques is how you get better analysis with AI.
Myth
I need to be using a particular tool to get good results.
Busted
If you see widely different results between models, that's a sign that you're relying too much on the model to make decisions.
Rephrase
I need to prescribe a sound analytical workflow to get good results.
Leo Hoar, PhD
Founder, UXR Institute
Leo founded the UXR Institute after years of working at the intersection of qualitative research and the kind of evidence skepticism that researchers routinely face in product and business settings. This mini-course grew out of the methods he developed for his full course on AI-assisted qualitative analysis, a workflow designed around the specific ways AI tools break when you hand them a corpus of interview data and ask them to "just find the themes." He's a practitioner first: the prompt in this course is the one he actually uses.
Work at your own pace
Nine video lessons you can complete in a single sitting or spread across a few days. Most people finish in under two hours, more if you practice with your own data alongside the lessons.
What a lesson looks like
Short explanations paired with real prompt examples and side-by-side output comparisons. Lesson 6 includes a full live demo of the prompt running on sample interview data, from alignment through the final pattern summary.
Practice as you go
The course includes sample transcripts from a fictitious autonomous vehicle rideshare project. Use those to follow along with the demo, or bring your own data. (No PII or confidential information in any AI tool.)
Do I need to know how to code or build AI systems?
No. The course uses standard chat interfaces like Claude and ChatGPT. You paste the prompt, paste your data in batches, and follow the steps. There's no programming, no APIs, and no technical setup beyond having an account with an AI chat tool.
Which AI tool should I use?
Any LLM chat interface works: Claude, ChatGPT, Gemini, or similar. The prompt is tool-agnostic. The course covers context window sizes for the common tools, which matters when working with larger datasets, and shows you how to adapt if your dataset is too big for one pass.
Is this just a prompt to copy?
The prompt is included and you're welcome to use it right away. But the lessons explain what each rule does so you can troubleshoot when something looks off, adapt the approach to your analytical question, and build new steps on top of the base workflow.
How long will this take?
The nine lessons run under two hours in total. If you work through the demo alongside the instructor and then run the prompt on your own data, plan for three to four hours. You can also stop after Lesson 6 and walk away with everything you need to run the prompt. The later lessons are about evaluation, customization, and a bonus on XML markup.
What if I want to go deeper than this mini-course?
This mini-course covers one structured prompt for analysis. If you want the full live curriculum, including rapid qualitative methods, coding and synthesis workflows, and cohort feedback from Leo, take a look at Using AI Responsibly for Faster and Deeper Insights.
