👋 Hey, I’m Nikki. Each week I write about UX research strategy, communicating impact, and using AI to do your best work. For more: Claude Skills Bundle | Claude Agents | AI Prompt Library | Team Training | AI Courses for UXRs
P.S. Paid subscribers get access to full archive, all content, a private Slack community, Substack lives, and a hub of templates, scripts, and mini-courses
Most researchers I talk to land in one of three camps:
You’ve never touched AI and you’re a little suspicious of the whole thing, which is fair
You use ChatGPT for the occasional email and you assume Claude is the same product with a different coat of paint
You’ve watched another researcher rave about Claude in a Slack channel and you’ve been meaning to try it for about four months.
This guide is for all three of you (and anyone else!), and I wrote it because most “intro to AI” pieces are useless for our work. They teach you to write a haiku or plan a trip to Lisbon, and then leave you to figure out the part that actually matters, which is how a tool that makes things up fits into a discipline built on evidence.
I’m going to teach you how Claude works, in a way that lets you predict when it will help you and when it will quietly hand you a fabricated quote with a straight face. I’ll show you how it compares to the other tools so you know what to use it for. I’ll walk you through getting set up safely, with the privacy settings that actually matter when you’re handling participant data. Then I’ll take you stage by stage through the whole research process, with copy-paste prompts and the exact spots where your judgment is the thing that can’t be skipped.
One thing before anything else, please be deliberate with participant data. Before you paste a transcript with someone’s name, employer, diagnosis, or salary into any AI tool, check your organization’s policy, strip the identifying details, and treat consent the way you already do everywhere else in your practice.
Part 1. What Claude is and how it works
You can’t use this tool well until you understand one strange fact about it. Claude doesn’t know things the way a database knows things. It predicts text.
When you type a question, Claude generates an answer one piece at a time, each time predicting the most likely next word based on patterns it learned from an enormous amount of writing. It’s doing this at a scale and speed that genuinely feels like thinking, and a lot of the time the output is excellent. but the underlying mechanism is always the same. It is producing the most plausible-sounding continuation of your text.
Pause for a second, because this explains every wonderful and dangerous thing AI will ever do for you. When you ask it to rewrite a clunky insight, “plausible continuation” gives you a sharp, well-structured insight, which is great. When you ask it to pull a quote from a transcript and the transcript doesn’t contain a clean one, “plausible continuation” gives you a quote that sounds exactly like something your participant would say, except they never said it.
Treat Claude like a brilliant, fast, eager junior researcher who has read almost everything, has no memory of your specific study, and has no personal stake in being right. Brilliant and fast means it can save you hours. No memory of your study means it only knows what you put in front of it. No stake in being right means it will hand you something confident and wrong as cheerfully as something confident and correct, and it’s your job to tell the difference. You would never let a junior put unverified quotes in a stakeholder deck. Same standard here.
A few behaviors follow from how the machine works, and understanding them is most of the battle.
It is confident whether or not it is correct
Claude produces fluent, well-organized text by default, so it sounds authoritative all the time. The tone is identical whether it’s quoting a fact it has seen a thousand times or inventing a statistic on the spot. There is no tell. It will not hedge unless you ask it to, and even then it hedges in the same smooth voice. For us, this means the fluency of an answer tells you nothing about its truth. A beautifully written synthesis can be built on a miscount and a perfectly phrased quote can be fictional. Your verification habits are the core of using this tool responsibly.
It wants to agree with you
Claude is trained to be helpful, and “helpful” gets nudged, over millions of training examples, toward “agreeable.” So if you walk in and say “I think the main problem is onboarding, what do you see in this data,” Claude has a pull toward finding evidence that you’re right. It’s simply predicting the response that best fits the framing you handed it, and you handed it a conclusion to support. No malice, no flattery, just the machine following your lead.
You already know why this is dangerous, because you spend your career fighting it in yourself and your stakeholders. It’s confirmation bias, and Claude can run it at high speed with total confidence. The fix is to never ask Claude to confirm, but to instead ask it to challenge.
It has a finite memory, and long chats go foggy
Claude reads and writes in chunks called tokens, where a token is roughly three quarters of a word. Every conversation has a fixed budget of tokens it can hold in mind at once, called the context window. Your prompt, every file you upload, and every reply all spend from that budget. A page of text is around 500 tokens, and a full study of transcripts can run to tens of thousands.
When a conversation gets very long, the window fills, and Claude starts losing track of things you told it earlier. The classic experience is a synthesis chat that was sharp an hour ago and is now contradicting itself and forgetting your research questions. That’s the context window overflowing. When a chat starts going AWOL, ask Claude to condense the important details into a handover document and start a new chat with the condensed context.
What it simply cannot do
Knowing the hard limits up front saves you a lot of frustration. Claude does not know what happened after its training cutoff unless you turn on web search, so it’s blind to last week’s news, your company’s internal updates, and anything proprietary it was never shown. It cannot do reliable arithmetic across a long document by eye, so any precise count needs it to run code. It cannot generate images, only read and analyze them. And it cannot watch a video or listen to an audio file directly, which matters for us, since a session recording has to be transcribed first by another tool before Claude can work with it. None of these are dealbreakers once you know them, they just tell you which jobs to hand it and which to prep first.
Part 2. Claude vs the other AI tools
You don’t have to pick one tool and marry it. Most researchers I know run two or three, each for what it’s best at.
ChatGPT is the one everyone’s heard of, and it’s the closest comparison. It’s strong, its voice mode is genuinely better than Claude’s, it generates images, and its quick web search feels snappy. Claude tends to write in a less robotic voice out of the box, holds very long documents together better, and works with your local files in a way ChatGPT doesn’t match yet. Claude for writing and heavy document work, ChatGPT for voice notes, images, and fast lookups.
Google Gemini is good, deeply tied into the Google ecosystem, strong on images, and handles long context well. If your whole world is Google Docs and Sheets, it’s worth a look.
Microsoft Copilot lives inside Office and is convenient if your org is all Word, Excel, and Teams, though for nuanced research writing I still reach for Claude.
NotebookLM deserves a specific mention for researchers, because it does one thing we care about really well. You give it a fixed set of sources, your transcripts, a report, a stack of papers, and it answers questions grounded only in those sources, with citations back to the exact passage. For desk research, literature reviews, or “what did participants say about X across these twelve files,” it’s excellent precisely because it stays inside your documents and shows citations.
Claude genuinely stands out at writing in your voice once you’ve shown it examples, reasoning over very long documents without losing the thread, working directly with the files on your computer, and running multi-step jobs through Cowork that other tools don’t really attempt. Use it for those, lean on the others for voice, images, grounded-citation Q&A, and in-repository tagging, and stop feeling like you have to be loyal to one logo.
Part 3. Getting set up the right way
Accounts, browser, and the desktop app
Go to claude.ai and sign up with your email. You can use Claude in your browser, which is fine for quick chatting, or as a desktop app you install on your Mac or Windows machine. For research work, install the desktop app, because it’s the only version that can see the files on your computer and the only place the more powerful modes live.
Plans, pricing, and managing usage
Free works in the browser with a limited number of messages a day and no Cowork, and it’s a fine way to test for a couple of weeks.
Pro, around 20 USD a month, gives you the smartest model, much higher limits, and the desktop features including Cowork. This is where most researchers land, and it’s the same price as ChatGPT Plus.
Max, at 100 to 200 USD a month, is for people running long, heavy jobs every day who keep hitting Pro’s ceiling. Team and Enterprise plans exist for organizations, and they matter for data handling, which I’ll come to in a moment.
Your usage is metered, and on the paid plans you get a generous but finite number of messages in a rolling window. If you hit the ceiling mid-task, you’ll be asked to wait a few hours or switch to a lighter model. The practical move is to save the most powerful model for the work that needs it and use a faster one for everyday drafting, which brings us to the models.
The models: Opus, Sonnet, Haiku
Claude comes in three sizes, and you pick from the selector in the composer.
Opus is the most capable, and it’s what you want for synthesis, nuanced insight writing, and anything where being right and subtle matters.
Sonnet is faster and lighter, a strong everyday workhorse for drafting, summarizing, and quick rewrites.
Haiku is the quickest and best for simple, high-volume tasks.
Reach for Opus when the thinking is hard or the stakes are high, drop to Sonnet when you want speed on routine work, and don’t overthink it beyond that.
Privacy and data settings, the part your legal team cares about
This is the section to read slowly, because it’s the one that determines whether you can responsibly put research data anywhere near this tool.
Two things to do:
First, in settings, find the controls for whether your conversations can be used to improve the models, and turn them off when you’re working with participant data.
Second, understand your plan. The consumer and Team and Enterprise plans handle data differently, with the business plans generally offering stronger guarantees, admin controls, and options some organizations require for sensitive data. I’m deliberately not quoting you exact policy language, because it changes and you should not run your ethics on my memory of a setting. Open your account’s privacy settings, read what they actually say today, and for anything regulated, get your organization’s answer in writing from whoever owns your data policy.
Turn training off for research data, prefer a business plan if your org handles sensitive participant information, and when in doubt, don’t put it in.
Teach it how you work, once
In settings there’s a space, often called “Instructions for Claude” where you describe yourself and how you want Claude to respond. Use it. Tell it you’re a UX researcher, that you write insights as key learning, why, and consequence, that you keep participant quotes verbatim, that you want it to flag anything unverified, and that you’d rather it challenged your assumptions than agreed with you. That short paragraph shapes every conversation from then on, so you stop repeating yourself.
If you want to go even deeper into this, you can create an about me markdown file which you give it access to in every project.
Part 4. The surfaces, modes, and building blocks
Claude is more than a chat bot, and the reason most people plateau is that the chat box is all they ever use. It’s really a set of pieces that fit together: modes, files, projects, artifacts, connectors, skills, and plugins. Here’s each one, what it actually is, why it matters for research, and how to use it. Learn these and you’ve learned most of what the so-called power users know.
Chat, Cowork, and Code
Chat is your fast thinking partner. You type, it answers, you go back and forth. Reach for Chat when the task is one contained piece of thinking that lives inside the conversation. Sharpening a single insight. Rewriting a finding three ways for three audiences. Drafting interview questions. Talking through whether your method fits your question. It’s quick, forgiving, works on free, and it’s where to spend your first week.
Cowork is your research assistant that does the work. You point it at a folder on your computer, describe a multi-step job in plain language, and it plans the steps, reads your files, writes new files, and checks in along the way. Reach for Cowork when the work spans several files, produces a real deliverable, and would normally cost you an afternoon. Tagging fourteen transcripts, building a synthesis matrix, cleaning a thousand survey open-ends, or drafting a readout from raw notes. This is the mode that gives researchers their time back, and the reason the desktop app is worth paying for.
Code is for engineers, with one rare exception for us. Claude Code runs in a developer’s terminal. If you don’t code, skip it. The only research case where it surfaces is genuinely heavy quantitative work, and even then you don’t open Code yourself, because Cowork can write and run the code behind the scenes and hand you the result. You can also use Claude Code to mock up visual representations of your findings like prototypes, but Cowork also has these capabilities. I recommend getting familiar with Cowork before moving to Code.
Uploading files, and what it can read
You can hand Claude files directly by dragging them in or clicking the attach button. It reads PDFs, Word documents, text files, CSVs and spreadsheets, slide decks, and images like charts, whiteboard photos, and screenshots. This is the everyday way you work with your own material: drop in a report and ask for a summary with page references, drop in a transcript and ask for themes, drop in a screenshot of a survey result and ask what stands out.
There’s a cap on how many files and how much total text you can attach to one chat, so for a big study you’ll want Projects or a Cowork folder rather than stuffing everything into a single message. And remember from Part 1 that it can’t process audio or video, so a recording has to be transcribed first by your notetaker or repository before Claude can touch it.
Projects: a saved workspace for a study
What it is. A Project is a saved workspace that bundles files, standing instructions, and a set of related chats in one place, so Claude carries the same context across every conversation inside it. Picture a room dedicated to one study, where everything you’ve pinned to the walls stays up between meetings.
Why it matters for you. A normal chat forgets everything the moment you close it, so you re-explain your study every time. A Project remembers, which means every chat inside it already knows your research questions, your transcripts, and how you like insights written.
What goes inside one. Two things. Project knowledge, the files every chat in the project can see, like your discussion guide, your anonymized transcripts, your screener, and your last report. And custom instructions, a standing brief for how Claude should behave in this project, like your insight format, your voice, and your guardrails.
How to set one up. Create a project, name it for the study, drag your files into project knowledge, write your insight format and guardrails into the instructions, then start your synthesis, insight, and readout chats inside it.
When to use which: a one-off question is a plain chat, a whole study you’ll return to over days is a Project, and a big multi-file job that produces deliverables is a Cowork folder. It’s one of the most useful features for researchers and one of the least used.
Artifacts: the panel where Claude builds things
What it is. When Claude makes something substantial, a document, a table, a chart, or a small interactive page, it opens that thing in a panel beside the chat called an Artifact, so what you’re building has its own space separate from the back-and-forth conversation.
Why it matters and how to use it. You can edit an artifact, ask for revisions in plain language (”make the recommendation sharper,” “turn this list into a table,” “add a column for the so-what”), and copy or export it when it’s ready. For us it’s where a readout draft lives while you refine it over several turns, or where an interactive, clickable view of your themes opens for a share-out instead of a flat slide.
Connectors: plug in the tools your research lives in
What it is. A Connector is a secure link between Claude and another app, like Granola, Google Drive, Notion, Slack, Gmail, or Zoom, that lets Claude read information straight from that tool so you don’t have to export and paste. You grant permission once, the same way you’d let one app connect to another.
Why it matters for you. Your research already lives in these tools, so a connector means Claude can pull a transcript from Drive or your notes from Granola without you copying anything across.
MCPs The open standard that makes connectors possible is called MCP. You’ll see the acronym thrown around, and you never have to touch it. A connector is just an MCP working quietly under the hood. Connecting a tool gives Claude permission to read from it, so think about what’s in there first, especially anything with participant data or confidential stakeholder threads, and follow the same data rules from Part 3.
Skills: your saved workflows
What it is. A Skill is a saved, named instruction set, a workflow you trigger by typing a slash and its name. Inside a skill is the same kind of thing you’d put in a really good prompt, your method spelled out, your format, your examples, sometimes a sequence of steps, packaged so you never have to retype it.
Why it matters for you. Most of us run the same handful of processes over and over: writing insights a certain way, pressure-testing a report, structuring a readout, building a screener. A skill captures one of those processes once and runs it on command, in your voice, the same way every time.
A research example. A /insight-writer skill might take raw findings and return them as key learning, why, and consequence, with a so-what check built in. A /screener-builder skill might turn a study brief into a recruitment screener with the disqualifying logic already worked out. Part 10 shows how to build your own library from the prompts that work for you.
If you’re looking for Claude Skills for user researchers, I highly recommend my 53-skill bundle based on 12 years of user research work.
Plugins: bundles you install
What it is. A Plugin is a bundle of skills and connectors packaged together, like a small app you install in one click, so you get a whole set of related workflows at once instead of assembling each by hand. You will also hear Plugins called Agents, which is an AI that plans and carries out multi-step work on its own, rather than answering one question at a time.
If you’re looking to install some UXR agents, check out my 7 highly trained user research agents.
Other surfaces
There’s a mobile app, good for capturing a thought or a quick ask on the go. There’s Claude in Chrome, a browsing agent that can move around websites and pull things up for you when there is no tool connection available (ex: Dropbox). And there’s Claude in Excel, for spreadsheet-heavy work. They’re all useful once you’ve got the fundamentals, and none of them is where you start.
The context folder that makes Cowork smart
One setup move pays off more than any other. Cowork gets dramatically sharper when the folder you point it at also holds context about how you work. Make a folder with three subfolders: an “about-me” folder with how you write, what a strong insight looks like to you, and the words you never want in a report; an “outputs” folder so Claude has a consistent place to save its work; and a “templates” folder with your best guide, readout, and insight format. Once that exists, your prompts get short and your outputs get sharp, because you’re not re-explaining yourself every time.
Part 5. How to talk to Claude
Most disappointing Claude sessions come down to a vague prompt. The instinct you already have, that a sloppy research question produces sloppy findings, is exactly right here. A prompt is a brief so you have to write it like one.
Here’s a skeleton I use for almost everything. You won’t need all of it every time, but when an output disappoints you, the fix is usually a missing piece.
Context: who I am, what study this is, what decision it feeds
Task: exactly what I want, as specific verbs
Source: the data, file, or example to work from
Format: how I want the answer structured
Guardrails: what to avoid, and what to do when unsureThat last line is the one researchers forget and need most. “If a verbatim quote isn’t an exact match, leave it out” and “if the reason isn’t in the data, tell me it’s a hypothesis” are the sentences that keep Claude honest.
Five habits that make every prompt better
Be specific, because “help me with my guide” gets you generic and “flag every leading question in this guide and rewrite it neutral” gets you usable.
Give it an example of your voice, because it learns faster from one good sample than from paragraphs of description, so paste a report you’re proud of and say “write like this.”
Say what you want rather than what you don’t, because “write it the way I’d explain it to a designer I like” beats “don’t be so formal.”
Build up rather than front-loading, since two sentences and then steering (”keep the exact words,” “cut it in half,” “add the workaround three people mentioned”) beats trying to write the perfect prompt in one go.
Start a fresh chat when one goes foggy, because a clean window thinks more clearly than a cluttered one.
A few advanced moves worth knowing
Give Claude a role, since “act as a skeptical research lead” or “act as a participant who’s skeptical of new tools” changes the whole frame of the answer.
Give it a rubric, since “score this guide against these five criteria” gets you something more rigorous than “is this good.”
Ask it to think step by step for hard analysis, which genuinely improves its reasoning on tangled problems.
Ask it to ask you questions first, with “before you answer, ask me anything you need,” which surfaces the gaps in your own brief.
And ask it to calibrate, with “how confident are you in that, and what would change your mind,” which is a fast way to find the soft spots in its reasoning.
Real prompts for real research tasks
Steal these, adapt the specifics, and notice the pattern so you can write your own.
Reviewing a discussion guide for leading questions, one of the safest high-value uses because there’s no participant data in it:
I'm attaching my discussion guide for a study on why mid-market
customers churn in their second month.
Do three things:
1. Flag every leading or double-barreled question and rewrite each
one to be neutral and open.
2. Point out where I'm asking people to predict the future or
explain someone else's behaviour, since they're unreliable at both.
3. Suggest three follow-up probes I'm missing in the moments where
the real story usually hides.
Leave my warm-up and closing as they are. No jargon.Coding open-ended survey responses, where the discipline matters as much as the speed:
Attached is a CSV of 480 responses to the open-ended question
"What almost stopped you from signing up?"
Group them into themes. For each theme give me:
- a short label
- the count and rough percentage
- two responses copied word for word, not paraphrased
- a one-line so-what
Then list anything that didn't fit a theme. Do not invent or tidy
quotes. If a verbatim isn't an exact match, leave it out. If you're
unsure of a count, say so rather than guessing.Spot-check a few verbatims against the CSV, and if exact numbers matter, ask Claude to count by running code rather than by eye.
Turning a flat finding into a real insight, the step where stakeholders are won or lost:
Here's a flat finding from my onboarding study:
"Users don't read the tooltips."
Rewrite it as a proper insight with three parts:
- the key learning: what we now understand about the behaviour
- the why: what's driving it, grounded in the evidence below
- the consequence: what it costs us if we ignore it
Use only what's supported by the quotes I'm pasting below. If the
"why" isn't in the data, label it a hypothesis instead of asserting
it as fact.
[paste 4 to 6 quotes]Triaging a vague stakeholder request into the real question:
A PM just sent me this: "Can you run a quick survey to find out if
users want dark mode?"
Act as a skeptical research lead. Tell me:
- the real decision hiding behind this request
- the assumptions baked into it
- whether a survey is even the right method here
- two sharper questions actually worth answering
Be direct and disagree with me where it's warranted. Red-team this request.Red-teaming your own interpretation, the antidote to the confirmation-bias problem and maybe the most valuable prompt here:
Here's my main insight and the three quotes I'm leaning on:
[paste]
Now argue the opposite. Using the transcripts I've attached, find the
disconfirming evidence, name the participants who don't fit the
pattern, and tell me honestly where I might be seeing this pattern
because I want to. Don't reassure me. Try to break it. Part 6. Where Claude fits across the research process
This is the part the generic guides skip, and it’s the part that matters. Claude threads through the whole research process, showing up at every stage with real help to offer and a checkpoint where your judgment stays non-negotiable. Here’s the honest map, stage by stage.
Intake and scoping
Claude is a sharp partner for turning a vague request into a real research question, surfacing the hidden assumptions, and naming the decision the research is meant to inform. Drop in the Slack message or the meeting notes and ask what decision this feeds and what’s being assumed. The stakeholder-triage prompt lives here.
You still have to decide whether the research is worth doing, and what’s really at stake politically, is yours. Claude can frame the question. It can’t read the room.
Study design
Claude can recommend methods, draft a discussion guide or screener, and pressure-test whether your method actually answers your question. The guide-review prompt is the workhorse. It’s good at catching the design mistakes we all make under time pressure, like asking people to predict their own future behavior.
The method choice is yours, because it depends on constraints, politics, and prior knowledge Claude doesn’t have. Use it to draft and stress-test, not to decide.
Recruitment and screening
Claude drafts screeners from a brief, writes recruitment and scheduling messages, and spots screener questions that telegraph the “right” answer to people angling to qualify. Low risk, high time saved, since there’s no participant data in it yet.
The criteria, the quotas, and who counts as a real user versus a professional survey-taker is still your expertise.
Fieldwork
Right after a session, Claude is useful for a fast debrief. Paste your rough notes and ask for the three things that surprised you, the moments worth revisiting, and the questions to sharpen before the next session. It keeps you analytical across a long week instead of saving all the thinking for the end.
Synthesis
Point it at a folder of transcripts and it will tag against your research questions, cluster the tags into candidate themes, pull representative quotes with their source, and draft a first-pass insight list. What used to be a two-day wall of sticky notes becomes a first pass you can react to in an afternoon.
Your checkpoint of deciding what the data means is the irreducible core of our job, and it stays with you. Claude can surface a pattern. Whether that pattern is signal or noise, whether it matters, what it implies for the product, that is judgment built on having sat in the room, and it does not transfer to a machine. Verify every quote against the source. Recount anything stated as a number. Treat the first-pass themes as raw material, not conclusions.
Insight writing and reporting
Claude is a strong writing partner for turning findings into structured insights, drafting an answer-first readout, and rewriting the same insight for different audiences. The insight prompt belongs here, and so does asking it to run a so-what check on a draft the way a skeptical stakeholder would.
The recommendation is yours. Claude can structure the argument and sharpen the language, but standing behind “here’s what we should do,” with your name on it, is the researcher’s responsibility.
Activation and the repository
After the readout, Claude helps keep insights alive, drafting the research newsletter, the Slack update, the one-pager for a leader who won’t read the full deck. It’s also useful for repository work, suggesting a tagging taxonomy, writing consistent summaries so past studies are findable, and helping you spot when a new question has already been answered by old research.
The pattern across every stage is the same. Claude is fastest at the drafting, the structuring, the first pass, and the tedious middle. You stay in charge of the judgment, the meaning, and the decision. Used that way, it protects your craft by clearing the busywork, leaving you more room for the part that was always yours.
Part 7. The full menu of UXR use cases
To make the breadth concrete, here’s a scannable catalog of things researchers actually use Claude for, grouped by phase.
Planning and design:
Turn a stakeholder request into a research question
Write a research brief
Choose a method, draft a discussion guide
Write a screener
Design survey questions
Draft a consent form starting point
Build a recruitment message and scheduling emails
Pressure-test a study plan
Desk and secondary research:
Summarize a long report with page references
Compare several vendor or competitor studies
Pull the relevant findings out of a stack of papers
Synthesize your own past research on a topic before you run anything new
Synthesis and analysis:
Tag transcripts against research questions
Cluster tags into themes
Build a synthesis matrix
Code open-ended survey responses
Draft an affinity structure
Surface patterns you might have skimmed past
Fnd the disconfirming evidence against your read
Insight and reporting
Turn findings into structured insights
Run a so-what check
Draft an answer-first readout
Write an executive summary
Rewrite an insight for a PM, a designer, and an exec
Draft speaker notes for a share-out
Modeling and artifacts:
Draft an evidence-based persona from your data
Sketch a dynamic journey map
Build an interactive chart or theme view for a readout
Operations and activation:
Design a repository tagging taxonomy
Write consistent study summaries
Draft the research newsletter or Slack update
Prep a stakeholder Q&A before a readout.
Part 8. Verifying Claude’s work
Verification is the price of using this tool in a discipline built on evidence, and it belongs in your habits rather than bolted on at the end. The good news is you already do this for junior colleagues, so it’s a habit you have, just pointed at a new source. Here’s the checklist I run before anything Claude produced goes anywhere with my name on it.
Check every quote against the source, by searching the transcript for the exact words. If you can’t find it verbatim, it doesn’t go in. Recount every number, and for anything that matters, have Claude count by running code rather than trusting an eyeballed total. Watch the magnitude words, because “some,” “many,” and “most” can drift, and “a few participants” quietly becoming “participants” changes the strength of your claim. Trace every claim back to evidence, and if a “why” or a recommendation isn’t grounded in the data, mark it as your hypothesis, not a finding. Turn on web search and double-check any external fact, date, name, or statistic. And read the whole thing once as yourself, asking whether you’d stake your credibility on it, because in the readout, you are.
The more consequential the claim and the harder it is for a reader to check, the more you verify. A reframed sentence in a brainstorm needs a glance. A participant quote on a slide in front of leadership needs to be confirmed word for word.
Part 9. The risks and ethics you own as a researcher
Every tool has tradeoffs, and the responsible thing is to know exactly where this one can hurt you. These matter more for us than for almost any other profession, because we handle people’s private words and we put our names on claims about what those words mean. Read this section twice.
It fabricates quotes and patterns that sound completely real. This is the headline risk. When Claude can’t find a clean quote, it can generate one that reads exactly like your participant, and when a pattern is thin, it can present one participant as a theme. It’s doing the only thing it ever does, producing plausible text, and a plausible quote is exactly its specialty, so a thin spot in your data gets filled with something that merely sounds right. Verify every generated quote against the transcript, and recount every number, before any of it reaches a slide. A fabricated quote in a readout is the kind of mistake that costs you years of credibility in one meeting.
It miscounts, and “some” becomes “all.” Claude is unreliable at precise counting across long text, and the magnitude drift is subtle. For anything where the number carries weight, have it count by running code, and check the strength words yourself.
It agrees with you, which is confirmation bias with a turbocharger. Never ask it to confirm your read, ask it to break it. Use the red-team prompt. The moment you’re most confident is the moment to invite the disconfirming evidence.
Confidentiality and consent are on you, not the tool. Anonymize before anything goes in: strip names, employers, exact job titles, locations, and any rare attribute that could identify a person, and replace them with labels like P04. You can even ask Claude to help anonymize a transcript, but then verify it caught everything, because a missed surname is still a breach. Check that your participants consented to their data being processed this way, and if your consent form predates AI tools, it almost certainly needs updating to cover them. And remember some data simply shouldn’t go into an external tool at all, regardless of how careful you are.
Data residency, retention, and IP are real questions for your organization, not for you to guess. Where your data is processed, how long it’s kept, and whether it can be used to improve models depend on your plan and your settings, covered in Part 3. Get the answer from your data owner in writing for anything regulated, rather than assuming.
The model carries bias, like any system trained on human writing. It can flatten an unusual viewpoint toward the average, or describe a group in stereotyped terms. For us that means watching that it isn’t smoothing away the very outlier or edge-case voice that often holds the most important insight. You are the check on that.
The deskilling worry is legitimate, and worth naming. If you let Claude do your synthesis and never look closely at your data, your analytical muscle weakens, and you become unable to catch the tool when it’s wrong. The way through is to use it for the first pass and the busywork while keeping your hands on the judgment, the meaning, and the verification. It’s a power tool, and the craft is still yours to keep sharp.
None of this is a reason to avoid the tool. It’s the operating manual. Researchers who internalize these risks use Claude with confidence, because they know exactly which parts to trust and which to check. The ones who get burned treated a fluent answer as a true one.
Part 10. Building your Skills library
I defined Skills back in Part 4, which are saved, named workflows you trigger with a slash. This is where you build your own, because that’s the move that takes Claude from a handy chatbot to a tool that runs your actual process on command.
The method is simple and it compounds. Once you’ve written a good prompt twice, stop rewriting it and turn it into a Skill. Keep a running note called prompts-that-worked, and every time something lands, the guide review that nailed it, the insight format that finally sounded like you, the so-what check that caught the weak finding, paste it in. After a couple of weeks you’ll have a personal library of your own best thinking, and the strongest of those become Skills you call by name: /insight-writer for your insight format, /so-what-checker to pressure-test a draft, /screener-builder to turn a brief into a screener.
The difference between a casual user and someone who genuinely relies on Claude usually comes down to this. The casual user retypes a paragraph of instructions every time. The committed one types /insight-writer and gets their exact method, in their voice, every time. Every good prompt you save is a piece of your craft you never have to rebuild, and that library will be worth more to you than any prompt pack you could buy.
Part 11. The directory
Prompt. The text you send Claude. Write it like a research brief, specific in, specific out.
Model. The version of Claude doing the thinking. Opus is the smartest, with faster, lighter Sonnet and Haiku underneath. Opus for synthesis and insight work, Sonnet for everyday drafting.
Token. The chunk Claude reads and writes in, roughly three quarters of a word. A page is about 500 tokens, and your limits and memory are measured in them.
Context window. The total Claude can hold in mind at once, in tokens. It’s what lets you load a whole study, and when it fills, Claude gets forgetful, so start a fresh chat.
Knowledge cutoff. The date the model’s training stops. It doesn’t know anything after that, or anything internal to your company, unless you provide it or turn on search.
Hallucination. When Claude states something false with full confidence, including invented quotes and miscounts.
Sycophancy. Claude’s trained tendency to agree with you.
Chat. The conversational mode where you type and it answers, for fast thinking, drafting, rewriting.
Cowork. The mode where Claude works for minutes to hours on a multi-step task, reading your files and producing deliverables, for synthesis, readouts, screeners, the afternoon-eating jobs.
Claude Code. Claude in a developer’s terminal. Mostly skippable, and Cowork can run code for you when needed.
Desktop app. Claude installed on your computer that sees your local files and unlocks Cowork, so this is the one to install.
Artifact. The side panel where Claude opens a document, table, chart, or small interactive thing it built.
Project. A persistent space that remembers files and instructions across many chats, a per-study workspace that keeps your method and voice on hand.
Project knowledge. The files and instructions every chat in a project can see like your guide, transcripts, and insight format.
Skill. A reusable instruction set you trigger by name.
Connector. A link between Claude and another tool like Granola, Drive, Notion, Slack, or Zoom.
MCP. The technical plumbing behind connectors.
Plugin. A bundle of skills and connectors, like a small app store.
Agent. A general term for an AI that plans and carries out multi-step work on its own.
Extended thinking. A setting that lets Claude reason longer before answering, worth turning on for genuinely hard analysis or design problems.
Web search. A tool you switch on so Claude can read the live internet, on for current facts, off when working strictly inside your own data.
Vision. Claude’s ability to read and analyze images you upload.
Custom instructions. A saved description of you and how you want Claude to respond.
Hallucination guardrail. A line in your prompt telling Claude what to do when unsure, like “leave it out” or “label it a hypothesis.”
Part 12. FAQs
Will it train on my participant data? It depends on your plan and settings, which is why part three matters. Turn off training in settings when handling research data, prefer a business plan for sensitive work, and get your org’s policy in writing for anything regulated.
Can I trust the quotes and counts? Not without checking. Verify every quote against the transcript and recount anything that matters, ideally by having Claude run code. This is the single most important habit in this guide.
Can it join my Zoom call or transcribe my recording? Not directly. It can’t listen to audio or watch video. Transcribe with your notetaker or repository first, then bring Claude the transcript.
Is it ethical to use AI on research data? It can be, with anonymization, consent that covers AI processing, and a plan that protects your data. It stops being ethical the moment you skip those or stop verifying what it produces.
Will it replace user researchers? It replaces the tedious middle of the work, not the judgment. Deciding what data means, choosing what to study, reading a room, and standing behind a recommendation are the job, and none of that transfers to a tool. The researchers who thrive use it to do more of the thinking, not less.
How is it different from the AI in my repository tool? Your repository’s AI is great for transcription and tagging where your data already lives. Claude is stronger for writing in your voice, reasoning across long documents, and multi-step jobs. Many researchers use both.
What’s the one thing to get right first? Verification. Everything else is upside. That one habit is what separates using this tool well from getting embarrassed by it.
Part 13. Your first two weeks
Don’t try ten random things. Ramp deliberately, so each step builds real intuition. Keep your data hygiene on the whole way by anonymizing before you paste, checking your settings, and checking your org’s rules.
Week one, stay in Chat and learn how it thinks. Set up your personalization paragraph first. Then on day one, paste three insights you’ve written and loved and ask Claude to rewrite three weaker ones in that voice. Day two, upload a long report and ask for a one-page summary with page references, then check two of them. Day three, run the guide-review prompt on a real discussion guide. Day four, run the stakeholder-triage prompt on a real request in your inbox. Day five, take one finding and ask for the insight three ways, for a PM, a designer, and an exec.
Week two, move to Cowork and Projects and feel the time come back. Build your context folder, and set up a Project for a current study with its files and your insight format. Point Cowork at a folder of anonymized transcripts and ask it to tag against your research questions and propose themes, then spend your time checking its work. Run the red-team prompt on your main insight. Ask it to draft an answer-first readout from your verified insights. And save the two prompts that worked best as the start of your Skills library.
How the pros use Claude: one command, a full readout
For everyone still with me, here’s what this can look like, without writing a line of code. I’d just wrapped a discovery study which included twelve transcripts, a pile of Granola notes, and a research plan with research goals and outcomes.
I didn’t write a long prompt. I’d already built a Skill that holds my entire readout process, so I typed one command and a single line of context.
In the few minutes I was gone, Claude read all twelve transcripts and tagged them against the research goals. It clustered the tags into candidate themes and pulled supporting quotes for each, with the participant ID and the spot in the transcript so I could check every one.
It ran my insight process, turning the strongest themes into structured insights, each with the key learning, the why, and the consequence. I’d built a so-what check into the same skill, so it pressure-tested every insight and flagged the two too thin to survive a stakeholder. Then it asked me the framing questions I should have been asking myself like which decision does this feed, who’s in the room, lead with the answer or build to it.
With my answers, it assembled an answer-first readout with the headline recommendation up top, the insights beneath it, the evidence underneath each one, every finding traceable to a quote, a recommended next step attached to each.
It didn’t decide what the findings meant, talk to the participants, or replace my read of which patterns were signal and which were noise. I verified every quote and every count, rewrote the insights in my own voice, and cut the threads that didn’t hold. What it gave me back was time to think, frame, and make important judgement calls.
Where to start
You can’t master this in a day, so don’t try.
Block 30 minutes this week.
Install the desktop app, set your personalization paragraph, and spend your first sessions in Chat teaching it your voice with examples of your own work. Run one of the prompts from part five on something real and sitting on your plate. When you’re ready for Cowork, build your context folder, set up a Project for a live study, and hand it one genuinely tedious job, a synthesis, a screener, a readout draft, and watch what comes back.
Verify everything, anonymize your data, mind your settings, and keep your judgment in the driver’s seat. Do that, and Claude stops being one more tool you feel vaguely guilty about ignoring, and becomes the research assistant you didn’t know you were allowed to have.
Open it tonight. I’ll be glad I talked you into it.
Stay curious,
Nikki
This is an absolutely incredible article - so incredibly thorough and helpful.
This is incredibly useful and I wish I’d had it when I first started using AI tools for user research! I will add a plus 1 for NotebookLM which was incredibly helpful for analyzing 110 of interview transcripts plus my own summaries for a meta-analysis of research over many months.