CASE STUDY · PRODUCT · UX RESEARCH · AI
A UX research workflow built on Claude: it turns Figma designs into self-recording prototypes and evidence-based reports, so eUp product teams can validate designs with real users before development.
Internal project — shown through screenshots only.
eUp’s product teams (Mazii, 4English…) ship new designs constantly. Yet a traditional usability round took about 4 weeks for 5 users: recruiting, waiting for developers to build a prototype, running each session, then rewatching hours of video to write the report.
By the time the report arrived, the design was already in a development sprint — the findings were right, but too late to act on. I redesigned the workflow around Claude and packaged it as an AI skill any product team can run.
Recruiting, scheduling, sessions, synthesis — a week each.
Designers had to wait for a working prototype from developers before testing.
By report time the design is in a sprint, and changes cost far more.
Rewatching hours of video, retyping notes, counting taps by hand.
Paid participants, staging, front-end effort for a build that exists only to be tested.
Because testing was so costly, many design decisions were made before any user data existed.
Design question · How Might WeHow might we let every product team validate designs with real users within the week, before development — without developer effort, and with results they can trust?
I broke a traditional round down week by week and role by role: designers waiting on developers, PMs waiting on the report, note-takers rewatching video. The bottleneck wasn’t the method but the time — results always landed after the design was locked.
The foundation is Jakob Nielsen’s discount usability (1989) and the 5-user rule (2000): test early, test with few people, fix, then test again. The problem was reframed — not replacing the method, but making it fast enough to run within a week, with no waiting on developers.
AI takes the labour-heavy steps: building the prototype from the design, logging, and writing the report. People keep the judgement calls: the PM sets hypotheses and criteria, the designer reviews the prototype, a moderator sits with each participant.
A full round fits into 6 actions and 3 message templates for Claude: invoking the skill, reviewing the prototype, and synthesising results.
The /prototype-usability-test skill turns a Figma design into an interactive web prototype that runs on a phone and records every tap, time on task and SEQ score (1–7). It provides a moderator link and a participant link, with one result code per person.
The report rates severity on Nielsen’s 0–4 scale, triangulates three layers of evidence (metrics – behaviour – quotes), places the current screen beside a proposed fix and states the confidence level of every finding.
The first round tested Mazii’s level-personalised home screen (29 Sep 2026): 5 participants, 8 tasks, 40 measured attempts. Only 1 in 5 could change their level from the home screen — a flaw in the core feature, caught before development; 5 of 5 found the home screen too dense. The report delivered 8 findings with fixes, a P1–P3 priority plan and acceptance criteria for round two.
To keep results trustworthy, the workflow follows 6 rules: lock criteria before the first session, discard pilot sessions, count by person (report 1/5, never as %), draw no CR/CTR conclusions, re-test with new people, and don’t conclude without enough participants. The skill has since validated a PM hypothesis for 4English Home and is taught to BO, PM and Marketing teams.
Onboarding website for new hires: a 5-level journey over time, summarised policies, a contact directory and an editing mode for HR.