사용 방법
- 설치 명령을 복사하고 자신의 에이전트 환경에 설치하기 전에 소스 패키지를 확인하세요.
- 필요한 도구, 인증 정보, 지원 환경은 원본 문서에서 확인하세요.
- 입력과 원하는 산출물을 설명하고 출처와 결과를 대조하세요.
npx skills add https://github.com/miratcan/claude-skills --skill persona-driven-ux-walkthrough시작 전 확인
- 설치에는 모델 구독, API 한도, 파일 접근 권한이 포함되지 않습니다.
- 별점 수와 설치 수는 활동 지표이며 독립적인 품질·보안 검증이 아닙니다.
원본 저장소의 제목, 예시, 코드, 표, 링크, 이미지를 유지해 표시합니다.
Persona-Driven UX Walkthrough Skill
When to Use
- User wants to test a web UI from a real user's perspective (not a technical audit)
- Building a skill/agent that browses a live site with Playwright and generates chain-of-thought output
- Designing a "persona simulation" feature for UX research
- User asks for "predictive user behaviour" in the context of UI/UX (not analytics/churn prediction)
Critical Disambiguation: "Predictive User Behaviour" ≠ Analytics
Common misunderstanding: "Predictive user behaviour agent" sounds like churn prediction, engagement analytics, or ML-based recommendations.
What users often actually want: A UX simulation skill that:
- Defines a user persona (age, tech experience, attention level)
- Browses a live site via Playwright
- Generates chain-of-thought "inner voice" output from that persona's perspective
- Reports UX scores + recommendations
Before assuming analytics: Ask what the agent is doing. If the answer involves "going to a page and narrating what the user sees/feels", it's a persona simulation, not analytics.
Skill Architecture
Single-Pass Skill (recommended starting point)
/ux-walkthrough <url>- Persona creation (interactive AskUserQuestion):
- Age range (18-25 / 26-35 / 36-50 / 50+)
- Tech experience (tech-native / moderate / tech-averse)
- Attention level (high / medium / low)
- Platform familiarity (first visit / heard of it / returning)
- Arrival context (how they got here: social link, Google, friend?)
- Optional: extra context
- Live browse: Playwright navigates to URL, captures screenshot + accessibility snapshot per page
- Per-page report:
- Inner voice (chain-of-thought narrative)
- Scores (comprehensibility, visual clarity, accessibility, first impression — each 1-10)
- UX recommendations
- Summary report:
- Top 3 critical findings
- Estimated drop-off page + reason
- Score averages
- Priority action list
Playwright Tools Used
browser_navigate — open URL
browser_snapshot — accessibility + DOM structure (aria labels, button text, etc.)
browser_take_screenshot — visual analysis
browser_click — persona clicks on elements
browser_press_key — scroll, tab navigationMental Model Calibration
The persona's inner voice should be calibrated by parameters:
| Parameter | Low | Medium | High |
|---|---|---|---|
| Tech experience | "What is this? Where do I click?" | "This looks like a login button" | "JWT auth, Bearer token header" |
| Attention span | 3 sec — if missed, gone | Scans page, looks for interest | Reads every element |
| Reading tolerance | Max 5 words | Headline + first sentence | Reads everything |
| Error tolerance | Leaves on first confusion | Tries once more | Debugs it |
Dopamine Tracking (dynamic, not static)
Dopamine level changes page by page, not set once:
- Increases: Beautiful visuals, sense of progress, clear CTAs
- Decreases: Long forms, confusing labels, broken interactions, language mismatches
- Arrival context sets initial dopamine: came from Instagram link → high curiosity; found via Google → neutral
Page Navigation Limit
Based on attention level:
- Low: 3-4 pages
- Medium: 5-6 pages
- High: 8-10 pages
Failed Attempts
| What I Tried | Why It Failed | Lesson Learned |
|---|---|---|
| Assuming "predictive user behaviour" = churn prediction/ML analytics | User meant persona-based UX simulation, not backend analytics | Ask "what does the agent do — does it visit a page and narrate?" before assuming analytics |
| Treating dopamine as a static persona parameter | Dopamine changes with each page interaction, not set once at start | Model dopamine as a running value that updates per page, not a fixed attribute |
| Skipping accessibility snapshot, using only screenshot | Screenshot shows visuals but not aria labels, contrast info, or focusable elements | Always combine screenshot (visual) + accessibility snapshot (structural) for complete analysis |
| Having agent jump to multiple parallel pages | Persona consistency breaks — each agent plays persona slightly differently | For persona simulation, use sequential single-agent approach, not parallel subagents |
Common Mistakes
- Don't: Build complex multi-agent pipeline immediately
Instead: Start with single-pass skill (one skill file, sequential pages) Why: Simpler to maintain, persona consistency preserved, context window manageable for typical 3-6 page flows
- Don't: Use static dopamine/mood level in persona
Instead: Track dopamine as a running state that changes with each page event Why: Real user engagement is dynamic — beautiful gallery raises it, confusing form drops it
- Don't: Only screenshot for analysis
Instead: Also take accessibility snapshot for DOM structure, aria labels, button targets Why: Snapshot reveals non-visual UX issues (missing labels, tiny touch targets, contrast ratios)
- Don't: Hard-code credentials in skill
Instead: Ask for credentials at skill start (optional), skip auth-gated pages if not provided Why: Skills should be usable without exposing secrets in the skill file
Output Format
Per-Page
### Page: <Name> (<url>)
**Inner Voice:**
> [Chain-of-thought narrative from persona perspective...]
**Scores:**
| Criterion | Score | Note |
|--------------------|-------|------|
| Comprehensibility | X/10 | ... |
| Visual clarity | X/10 | ... |
| Accessibility | X/10 | ... |
| First impression | X/10 | ... |
**Dopamine state:** [rising/neutral/falling] (X/10)
**Recommendations:**
- [Specific fix]Summary
## Summary
**Persona**: [brief description]
**Estimated drop-off**: [Page X, reason]
**Score averages**: Comprehensibility X | Clarity X | Accessibility X
### Top 3 Findings
1. [Critical issue]
2. [Critical issue]
3. [Critical issue]
### Priority Actions
- [Actionable fix with page reference]See Also
- Examples — Real walkthrough examples
- Troubleshooting — Playwright interaction issues discovered
