Guides user research and research ops. Use when running interviews, usability tests, surveys, or synthesis to de-risk product decisions; not experiments or review mining.
Independent GitHub snapshot · Automated classification · Original documentation governs
name: software-ux-research
description: "Guides user research and research ops. Use when running interviews, usability tests, surveys, or synthesis to de-risk product decisions; not experiments or review mining."
compatibility: Portable core. Works on Claude Code and Codex.
version: "1.3"
last_validated: 2026-08-10
Software UX Research
Use this skill to reduce product and design risk with evidence. It owns research method choice, study design, findings synthesis, and research operations. It does not own UI implementation.
Quick Reference
Need
Default
Output
discovery and JTBD
semi-structured interviews with a predeclared segment, saturation, and budget stop rule
opportunity brief
usability evaluation
moderated tests covering each critical task and user group; iterate in small rounds
findings report with severity
quantification after qual insight
survey or analytics review
segment or pattern readout
causal change validation
controlled experiment or staged rollout
experiment brief
research ops and repository design
lightweight intake, taxonomy, and consent model
research-ops recommendation
accessibility or low-digital-literacy research
moderated sessions with adapted materials
risk and inclusion report
When to Use This Skill
Use this skill when the main question is:
what user problem matters and for whom
whether a concept, flow, or prototype is understandable and usable
which research method is appropriate
how to design a study and synthesize findings
how to run research ops, repository, and consent workflows
choose the smallest method mix that can answer the question
use qual for motives and friction, quant for scale and segmentation
treat synthetic participants as hypothesis generation only
require confidence level and evidence trail in every output
current standards and regulatory claims must be verified before final advice
Quality Lens
Choose the evidence layers the decision needs: task success and friction for a flow fix; emotion and meaning when studying trust, retention, or adoption. State which layers were studied and limit conclusions to them. Load references/consumer-experience-quality.md for the four-layer lens and its methods; task completion alone does not establish a retention benefit.
blocker identification, accessibility, and readiness
growth
retention, friction, and segment behavior
maturity
optimization, simplification, or feature retirement
Verification Checklist
Before delivering any research output:
Decision the study was designed to unblock is named explicitly
Method justified: weaker alternatives were considered and rejected with reasons
Participants match the target segment — not convenience, panel-only, or CS rolodex
Sample rationale names tasks, segments, risk, excluded populations, numeric stop threshold, consecutive evaluation window, maximum sample/budget, and action at the cap; experiments are power-calculated
Confidence level and evidence trail stated in the output
Synthetic participants labeled as hypothesis generation only — not cited as evidence
AI-assisted analysis plan predeclares the tolerated disagreement by severity, audit-batch size, consecutive passing batches, maximum audit size, and the human-recoding/escalation action if the cap is reached
Consent obtained; recordings, transcripts, and participant identity stored separately
EU/UK participant data: DPA in place before sending to AI-processing vendor; check whether EU AI Act high-risk (Annex III) deployer obligations apply and take their current application dates from qualified EU regulatory counsel or the Official Journal text, not from memory
Disconfirming evidence documented, not only confirming clips
Agentic products: study ran multi-turn, exercised at least one interruption, and included seeded incorrect outputs if trust was measured
Research Ops Rules
capture the decision, audience, segment, and evidence links in intake
use one taxonomy across studies and atomic insights
separate participant identity from notes and recordings
redact broad-share artifacts
let non-researchers run only templated studies with review guardrails
Prefer ISO, W3C, regulator, and primary-method sources over summaries.
AI and Accessibility Notes
For AI-powered product research (the thing being studied is AI-driven):
test trust calibration, failure recovery, explainability, tool-use disclosure, and approval gating
separate wrong output from unclear output and non-recoverable failure
run multi-turn sessions for agentic products — single-turn studies miss most of the failure surface
test steering explicitly: users change their mind mid-task, and addition/revision/retraction fail differently
measure trust calibration against seeded incorrect outputs; an all-correct study cannot distinguish good judgment from blind acceptance
For AI in the research workflow (synthesis tools, AI moderators, synthetic users):
treat synthetic users as hypothesis generation only (NN/g position), never as evidence
start analysis from a human-coded seed sample, then let AI extend; audit every tag class plus low-confidence and random items against the predeclared disagreement rule
AI moderators are appropriate only when the protocol is structured enough for a junior human to follow
inventory every AI tool that processes participant data (vendor, data sent, DPA, AI Act classification — see the EU/UK checklist item above and references/ai-in-research.md#eu-ai-act)
recruit assistive-technology users when accessibility is in scope
distinguish accessibility usability findings from formal conformance findings
Known Traps
Starting with a preferred method before naming the actual decision the study needs to unblock.
Recruiting convenience participants whose context, literacy, or workflow is too far from the target segment.
Treating generated summaries, AI note clustering, or synthetic participants as evidence instead of support material.
Mixing discovery, usability, and causal-validation questions into one study and getting ambiguous output from all three.
Reporting severity or confidence without tying it to sample quality, task coverage, and evidence strength.
Storing recordings, transcripts, and participant identity with weaker controls than the sensitivity of the study requires.
Sending EU/UK participant recordings to a non-EU AI vendor (Dovetail, Marvin, Looppanel, or any foundation-model-backed service) without a current DPA and explicit AI processing disclosure in consent — Chapter V GDPR transfer rules apply.
Recruiting only from professional research panels (Prolific, UserTesting panel) for behavior studies, then generalising to product users — panel respondents are experienced participants whose behavior systematically diverges from first-time real users.
references/westrum-culture-measurement.md — worked example of a validated seven-item team-culture survey construct (Westrum typology, scoring, provenance limits); pairs with foundations-team-theory
references/ab-testing-implementation.md — experiment vs research decision, readout checks, maturity levels; experiment mechanics route to marketing-product-analytics
When prior decisions or pitfalls are relevant, consult learnings.consolidated.md if present; use learnings.md only for needed history or as the available fallback. Otherwise skip both.
After applying it, if you encountered a pattern worth remembering, a mistake worth preventing, or a domain fact that surprised you, append one dated bullet to learnings.md via agents-skills-feedback-loop/scripts/append_learning.py. Do not modify SKILL.md itself.