REMORSERemorsearch
UX QA
An Agent Skill for Claude Code that improves the UX of an existing product. Claude examines, tests, researches, fixes and verifies, and every finding carries its evidence.
> /remorsearch-ux-qa Improve the checkout UX URL: https://shop.example.test ● Examine product, docs, recorded decisions ● Test persona: first-time buyer (synthetic) [observed] Card form resets after a failed payment [inferred] Buyers may drop off at re-entry [unknown] Share of mobile buyers (no data) ● Research 4 public sources · 2 coverage gaps ● Fix keep valid fields (reversible) ● Verify replay passed · no regression >
Five steps. No step skipped.
Give Claude a URL or a repository, and ask for an improvement. A problem list, personas or a research brief are optional.
Examine
Claude reads the product and the recorded decisions of its developers, so intentional choices stay intact.
Test
Baseline and persona QA: task completion, input preservation, error recovery, keyboard use, KWCAG and WCAG 2.2.
Research
Real user experiences from permitted public sources, with provenance and personal data removed.
Fix
Scoped, reversible fixes. Screen changes go to Figma first, with the file, node and version on record.
Verify
A replay on the current build, an adjacent regression check and an intent check. Only then is it done.
The agent picks the components.
You do not select a mode. A research-only or review-only request does not change the product.
| Component | Function | Requirement |
|---|---|---|
| research | Finds user experiences, workarounds and counterexamples in communities, reviews and official sources. | A permitted source. YouTube comments need your own YouTube Data API key. |
| persona-qa | Tests the real product with synthetic test conditions. | Access to the product and safe test data. The live web driver needs Node.js 22 and Playwright. |
| ergonomics (experimental) | Screens reach, touch, perception, timing and layout with simulated profiles. | Applicable scenarios. |
| loop | Records scoped fixes, replays the current build and checks for regressions. | The browser and Figma tools of the host agent. |
One folder. Done.
Python 3.10 or later on macOS, Linux or WSL. The Python tools use only the standard library.
git clone https://github.com/remorser58/remorsearch.git
mkdir -p ~/.claude/skills
cp -R remorsearch/skills/remorsearch-ux-qa ~/.claude/skills/
# then, in Claude Code
/remorsearch-ux-qa Improve our app UX. URL: https://app.example.testEvidence rules
- Each statement has one label: observed, inferred or unknown.
- Synthetic personas make test conditions only. They do not show the preferences of real users.
- A community signal is a hypothesis until Claude confirms it on the real product.
- Deployment, publication and changes to real accounts need a separate permission.
Responsible research
- Reads only public pages. Stops at logins, paywalls, CAPTCHAs, bot blocks, robots.txt rules and AI or TDM opt-outs.
- Does not disguise itself. One page at a time per host, 8 seconds or more apart.
- Removes phone numbers, emails, ID, card and account numbers and IP addresses before Claude reads the text.
- Korean pack: 27 Korean channels with access routes and terms status.