user-research-basics
Skill2 starsChanged 11 days agoMITView on GitHub
Use whenever someone needs to find out what people do, need or struggle with, in any words, length or language - "should we research this", "we need to understand our users", "write a screener or interview guide", "why did people cancel or switch", "is this survey okay", "make sense of these notes or support tickets", "what do these test results mean", "create personas", "can we ship this test result", "talk to stakeholders before we start". Covers scoping, cheapest method, small-study plans, neutral questions, bias and consent, synthesis, finding priority, personas from evidence, experiment hygiene. Not for judging a design (design-critique), form fields (form-design) or page copy.
In this repository25 other files it gives its agents.
- agent-interface-designSkill
- brand-identitySkill
- chart-designSkill
- color-paletteSkill
- data-tablesSkill
- design-critiqueSkill
- design-system-builderSkill
- form-designSkill
- generated-ui-restyleSkill
- interface-behaviorSkill
- journey-and-service-mappingSkill
- layout-structureSkill
- practice-and-assessmentSkill
- presentation-designSkill
- pricing-and-packagingSkill
- responsive-layoutSkill
- sales-conversationsSkill
- showing-workSkill
- spacing-and-groupingSkill
- teaching-interfacesSkill
- typeface-selectionSkill
- typesettingSkill
- ui-motionSkill
- ui-polish-passSkill
- visual-hierarchySkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 80
Every number, quote and participant came from the user's material.
- Distinctive: line 82
No question asks for wants, likes or predictions.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
User research basics
Research here means a clear question tied to a decision, the right few people, neutral asking, honest counting and a short written trail. This skill plans studies and analyzes material the user supplies. It cannot recruit anyone or run a session, so it always says what a human must do. The default failures are a survey asking what people want, five hundred respondents instead of five conversations, invented personas, percentages from a handful of people, and findings that jump to features.
Do not produce
- A participant, quote, count, percentage, persona detail or "users said" line that is not in the supplied material. Missing evidence becomes a labeled hypothesis or
[to collect]. - A survey or interview asking what people want, whether they like it, or whether they would use it.
- Personas, journey maps or segments with no source data. Offer labeled assumptions plus a plan to test them.
- Percentages or "significant" from fewer than about thirty people; say "3 of 7".
- Solutions inside findings, or findings that name no source.
- A long academic plan, or "research is needed" with no plan.
- Focus groups as a method; a friend, colleague or enthusiast as the whole sample.
Missing facts
Deliver on stated assumptions, in one line ("assuming a consumer app and no existing interviews"). Ask once, with a default, only if interactive and cheap; never reply with only questions. Do not invent traffic, sample sizes, budgets, dates or who attended.
First decision: is research warranted?
Name the decision, who will act on it and what a wrong guess costs in months of work. Then check what already exists: tickets, chat logs, sales notes, search queries, reviews, analytics, earlier studies. Default: use existing evidence first. Research only the expensive assumptions; skip it for a throwaway prototype or a choice that is cheap to reverse. If a person insists on a method that cannot answer the question, say what it can and cannot tell them and propose the cheaper fit.
Pick the method from the question
Phrase the question so you can tell when it is answered: name the output (a list of X, a comparison of Y). Add the method and the result that would change the plan. Smallest versions to start from:
- What are people doing and why, problem unknown: five or so conversations about recent real behavior, ideally where it happens.
- A candidate design or flow: three to five people doing real tasks, one at a time, nobody helping. Section "Watching people" in design-critique's checks, if installed, is the quick version.
- What happened, how much, where: analytics or existing records first, then people for why.
- Which of two wins: a live comparison, only with enough traffic and one pre-stated metric.
- How many, once the themes are known from conversations: a short survey of the right people. Opinion polling and focus groups answer neither "what do they do" nor "will they buy".
Participant counts are rough guides, never statistics. Three to five for a quick check of one kind of user; five to eight per kind when there are several kinds or a fix-and-retest cycle. Stop when new sessions stop changing your conclusions; add a few when they still do. Details in plan, screener and asking.
Participants
One test: would this person do the task this week? Screen on what they did recently, not on skill or opinions, and bury the target activity among others so the topic cannot be guessed. Keep the screener short, exclude people who guess to win the incentive. Say who the sample leaves out.
Asking
Ask for stories about the last real time, not opinions or predictions. Open questions, "tell me more", silence. Never mention the product or sponsor before the materials show it. Get consent for recording and for any unseen observers; end a session if the person is driving or distressed. The designer of the thing is the worst facilitator of its test.
Synthesis of supplied material
- Code every note to its source; separate what happened from what it means.
- Group bottom-up, then name each group as a need, barrier, trigger or word people use.
- Count as n of N, with the sample's source and bias ("7 interviews; support tickets come from people who write in").
- Chain: observation, pattern, insight, a decision to make. Rank by evidence strength and by importance to the goal; mark single-source items as leads. Set aside outliers who could not be users and say why.
- End with what is still unknown and the cheapest next check.
Depth and personas in synthesis, findings and evidence.
Usability findings
Each problem: task, how many of N hit it, what happened, severity (blocks, slows, annoys). Weight by how central the task is. Preference is not a finding. Fix by renaming, moving or removing before adding explanation; retest with new people. Frequency bands are a sorting aid at this size.
Numbers
Analytics says what and where, people say why. An experiment needs one metric chosen first, a run covering whole weeks, no early peeking and enough traffic; a few days and hundreds of visitors is a lead, not a result. See stakeholders and experiments.
When not to use
- Judging a design someone shows you: design-critique if installed. Form field design: form-design if installed.
- Statistical modeling, market sizing, population-scale panels: a short recommendation, and a specialist runs it. A survey of your own users or members stays in scope.
- Children, patients, regulated or sensitive topics, or politically charged organizations: recommend a trained researcher and give a plan outline only.
- The user's constraints, schedule and stated method win when they can work; flag the risk in one line.
Output discipline
Deliver the requested artifact first (plan, screener, guide, synthesis, ranked list) with no preamble, because the reader came for the work. Give exactly the count asked for. Honor any length cap and count rather than assert. After the work, at most two one-line notes (assumption, what a human must do). If the existing plan, screener or guide already works, say so and change at most a word.
Quick checks
- The decision and the result that would change it are named.
- Every number, quote and participant came from the user's material.
- Counts are n of N with the sample's bias; no percentages from small groups.
- No question asks for wants, likes or predictions; no leading words.
- What a human must do (recruit, consent, run, record) is stated.
References
- references/plan-screener-asking.md: load for a plan, screener, interview guide, leading-question rewrite, bias or consent text.
- references/synthesis-findings-evidence.md: load for notes or tickets to analyze, usability findings, personas, scenarios, or reviewing existing evidence.
- references/surveys-measures-inclusion.md: load to write or review a survey or questionnaire, plan a sample, report task success, time or ratings from a test, or recruit, pay and consent people.
- references/purchase-and-switch-interviews.md: load to interview people who recently bought, switched, cancelled or walked away, or to code and count what they say.
- references/running-sessions.md: load to script or review how a moderator runs an interview or task test, or to fix sessions that went badly.
- references/testing-designs-and-prototypes.md: load to plan a test of a design, prototype or pilot: what to show, task wording, value questions, how far a small round can be trusted.
- references/stakeholders-experiments.md: load for mapping stakeholders, a success-and-scope note, split tests or analytics interpretation.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

