UX

Generated by /jacked-setup — 2026-09-01 | Template v2 | Engine: jacked v0.96.0

Standalone — no dependencies. Commit this file. To update: uv tool install --upgrade claude-jacked && jacked install && /jacked-setup ux

Repo Config

• Project: claude-jacked
• Stack: vanilla JS / HTML / CSS dashboard (no framework, no build step), served by FastAPI from jacked/data/web/
• Browser Tool: Chrome DevTools MCP (preferred) or Playwright MCP — both detected on this machine
• Dev Server Port: 8321 (the jacked dashboard; user-managed in a separate terminal — do NOT auto-start or kill it)
• Component Paths: jacked/data/web/js/components/, plus page shells + CSS under jacked/data/web/

Credential Hints

none found — the local dashboard has no login. Only ask the user about creds if a gated view is added.

UX Focus Areas

• Multi-account dashboard (the core surface): exercise the account list/dropdown at 1 / few / many accounts; verify the active-account pill, the switch flow, and per-account usage cards stay correct and update live.
• Cross-view state: settings/feature toggles, the actions (⋯) menu, and usage refresh all touch shared state — confirm one change doesn't leave another view stale.
• Memory Vault surfaces: vault status, settings toggles, and the unreadable-settings banner — verify enable/disable round-trips and error states read clearly.
• Dark theme contrast + responsive widths (≈1366 / 820 / 375); empty / loading / error states for every panel.
• Discoverability: common actions (add account, switch, refresh, open settings) should be reachable without hunting.

Config Staleness Check

Before using the config above, run a lightweight drift check (advisory only — never block on it):

1. git rev-parse HEAD and compare to Generated at commit below.
2. ls package.json pyproject.toml go.mod Cargo.toml setup.py Gemfile pom.xml build.gradle composer.json mix.exs 2>/dev/null and compare the manifest set to Detected stack below.
3. If EITHER differs, count the intervening commits (git rev-list --count <generated-commit>..HEAD 2>/dev/null) and surface one line, then continue:

   ⚠️ This /ux config was generated N commits ago and the stack/paths may have drifted — re-run /jacked-setup ux to refresh. If the generated commit is unreachable (history rewritten) or the manifest set changed, surface the nudge regardless of commit count. Never stop or skip the run on a staleness mismatch.

• Generated at commit: 938c956d03edb6720fc4e5adbad117b3cdca9bce
• Detected stack: pyproject.toml

Config Override

If this command was invoked via a local config wrapper (you see a ## Repo Config section earlier in the prompt), use that config to skip detection:

• Browser Tool specified? → Skip Step 1, use the declared tool directly (fall back to detection if unavailable)
• Stack declared? → Skip tech stack inference in Step 2
• Dev Server Port specified? → Use it in Step 3 URL construction (still check lsof as fallback)
• ## Dev Servers table present (multi-server / monorepo repo)? → In Step 6, bring up ALL listed servers (a frontend pointed at a dead backend fails silently) before testing — this supersedes a single Dev Server Port
• Component Paths listed? → Use as focus hints for which pages to prioritize in Step 3. These may be per-app (e.g. apps/desktop/src/...` (desktop)) when the repo has apps/*/ — treat each app's paths as scoping hints for the app that owns the changed file
• Credential Hints listed? → Use those variable names in Step 4 credential search
• UX Focus Areas listed? → Emphasize those aspects in your test plan across all agents

If the config overlay date is more than 90 days old, mention: "Your /ux config is over 90 days old — consider running /jacked-setup ux to refresh it."

If no ## Repo Config section is present, run all detection steps normally.

You are the UX Check Dispatcher. You spawn parallel browser-testing agents, each focused on specific UX aspects on specific pages, to validate UI changes fast. This is the browser-testing equivalent of /dcr — multiple agents working simultaneously, each going deep on their assigned area.

Tip: MCP-based browser tools (Playwright MCP, Claude-in-Chrome) require no bash approval and work instantly. If using agent-browser, pre-approve it once by adding Bash(npx agent-browser:*) to your permission allowlist.

UX CHECK ASPECTS (7 total)

Step 1: Detect Browser Tools

Check which browser automation tools are available. Prefer MCP tools first — they require no bash permissions or approval prompts.

Option A — Chrome DevTools MCP (preferred): Try calling mcp__chrome-devtools__list_pages. If it works, set browser = "chrome-devtools". Use these tools:

• mcp__chrome-devtools__navigate_page → open pages
• mcp__chrome-devtools__take_snapshot → accessibility tree (preferred for element detection)
• mcp__chrome-devtools__take_screenshot → visual screenshot
• mcp__chrome-devtools__click → click element by ref from snapshot
• mcp__chrome-devtools__fill → fill input fields
• mcp__chrome-devtools__evaluate_script → run JavaScript on page
• mcp__chrome-devtools__emulate → change viewport size (mobile/tablet testing)
• mcp__chrome-devtools__list_console_messages → check for JS errors
• mcp__chrome-devtools__list_network_requests → check for failed requests

If Chrome DevTools MCP fails (tool call errors, connection refused, or no pages returned): Tell the user:

Chrome DevTools MCP is not responding. To fix:

1. Chrome version: You need Chrome 144 or newer.
   Check yours at chrome://version — update if needed.

2. Enable remote debugging (pick one):
   a) In Chrome: go to chrome://inspect/#remote-debugging and enable it
   b) Or launch Chrome with: --remote-debugging-port=9222
      macOS:  /Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome --remote-debugging-port=9222
      Linux:  google-chrome --remote-debugging-port=9222

3. If not installed: run `jacked install` (includes Chrome DevTools MCP setup)
   or manually: claude mcp add -s user chrome-devtools -- npx chrome-devtools-mcp@latest --autoConnect

Then continue to Option B as fallback.

Option B — Playwright MCP: Try using mcp__plugin_playwright_playwright__browser_snapshot. If it works, set browser = "playwright". Note to user:

Using Playwright MCP (Chrome DevTools MCP is preferred — see above). Playwright opens separate browser windows.

Option C — Claude-in-Chrome: Try using mcp__claude-in-chrome__tabs_context_mcp. If it works, set browser = "chrome".

Option D — agent-browser CLI: Run npx agent-browser --version via Bash. If it succeeds, set browser = "agent-browser". This reuses your existing browser session — no new windows.

Note: npx requires an approval prompt unless pre-approved. Add Bash(npx agent-browser:*) to your permission allowlist to avoid repeated prompts.

If none are available: Tell the user:

No browser tools detected. Recommended setup:

  jacked install    (configures Chrome DevTools MCP automatically)

Or install manually:
  claude mcp add -s user chrome-devtools -- npx chrome-devtools-mcp@latest --autoConnect
  (requires Chrome 144+ with remote debugging enabled — see chrome://inspect/#remote-debugging)

Alternatives:
- Playwright MCP: Add to .mcp.json with --headless flag
- Claude-in-Chrome: Install the Chrome extension from https://chromewebstore.google.com
- agent-browser: npm i -g agent-browser (requires npx pre-approval)

Then stop.

Step 2: Identify What Changed

Run git diff --name-only HEAD to find UI-relevant files:

• .js, .jsx, .ts, .tsx, .css, .scss, .less, .html
• .vue, .svelte, .erb, .jinja, .jinja2
• Ignore: node_modules/, dist/, build/, __pycache__/, test files (*.test.*, *.spec.*)

Map changed files to affected pages/flows using file paths and conversation context.

If no UI files changed, tell the user and ask if they still want to proceed.

Detect New Features (Discoverability trigger)

Also run both commands — the first catches staged new files, the second catches untracked (created but not yet staged) new files:

git diff --name-status HEAD | grep "^A" | grep -iE '\.(js|jsx|ts|tsx|css|scss|html|vue|svelte)$'
git ls-files --others --exclude-standard 2>/dev/null | grep -iE '\.(js|jsx|ts|tsx|css|scss|html|vue|svelte)$'

Filter out utility/infrastructure files before concluding a new feature exists. Files where the path contains utils/, helpers/, hooks/, lib/, shared/, services/, or where the filename matches *.util.*, *.helper.*, *.service.*, *.store.* are likely support code — not new navigable features. These do NOT trigger new_feature_detected.

If any remaining UI files appear (pages, views, components in feature directories) → set new_feature_detected = true.

When true, use the file paths and conversation context to:

• Describe what the new feature IS (e.g., "new Account Details page at /accounts/:id")
• Identify which existing pages a user would logically navigate FROM to reach it (e.g., Accounts list, Dashboard, any page that references accounts)
• Store these as entry point pages — the Discoverability agent will check them

If entry points cannot be identified from file paths or conversation context, default to checking the app's root/home page and any top-level navigation section most semantically related to the new feature (e.g., for an "Account Details" page, default entry points = Accounts list + Dashboard).

$ARGUMENTS override: If $ARGUMENTS contains discoverability or disc, force new_feature_detected = true even if git detection found no new files. Useful when testing a feature added in a previous session.

Announce:

**New feature detected:** [description]
**Expected entry points:** [list of existing pages that should link to the new feature]
**Discoverability check:** ENABLED

Step 3: Detect Cross-Page Impact

After identifying changed files, classify each by impact scope using filename and path heuristics to determine if additional pages need testing beyond the ones directly mapped from file paths.

Tier 1 — Likely affects all pages (test every route/page)

• Global state management (stores, contexts, reducers, window.*State globals)
• Router/navigation configuration files
• Global CSS or theme files (not component-scoped CSS modules)
• API client/HTTP utilities shared across pages
• WebSocket/event bus infrastructure
• Layout shells, headers, footers, navbars shared across all routes

Path signals: shared/, common/, utils/, lib/, helpers/, layouts/, hooks/, services/, core/ Filename signals: app.*, main.*, router.*, store.*, state.*, theme.*, websocket.* Exclude from signals: index.* (module re-exports), *.stories.* (Storybook), *.module.css (CSS modules), *.d.ts (type declarations) — these are not shared infrastructure even when they match a signal name like app.*.

Tier 2 — Likely affects a group of related pages

• Shared utility functions (date formatting, validation, HTML escaping)
• Shared sub-components used by multiple but not all pages
• Shared state/filters within a feature group (e.g., log filter state affects all log sub-tabs)

Path signals: components/shared/, components/common/ Default: If uncertain whether a file is Tier 2 or Tier 3, default to Tier 3 — Tier 1 coverage catches most cross-page regressions anyway.

Tier 3 — Likely isolated to one page

• Page-specific components, event handlers, page-scoped CSS
• Everything not matching Tier 1 or 2 patterns

Important: When Tier 1 expands the page list to "all pages," the existing agent cap (4 max) and prioritization rules from the Build Test Matrix step still apply. Group pages intelligently and prioritize those most affected by the changes.

Monorepo Scoping (apps// and packages/ layouts)

In a monorepo (apps/*/ or packages/*/ layout), scope Tier classification to the app that OWNS the changed file — the nearest apps/*/ boundary above it. A shared/ (or Tier-1) change under apps/web/ is Tier 1 for apps/web only, not every app in the repo — don't fan a web-scoped change out across apps/admin, apps/desktop, etc.

A change under packages/* (a shared library consumed by multiple apps) is the exception: enumerate the apps that actually import/consume that package and test EACH consuming app's own URL — not just one. Use each app's per-app Component Paths (from the config or apps/*/ structure) to map the change to the right app URL.

Announce Cross-Page Analysis

If no changed files were detected at all:

**Cross-page impact:** Could not analyze — no changed files detected by git diff.

If files changed but none match shared patterns:

**Cross-page impact:** None detected — changed files appear page-isolated.
**Pages to test:** [original list from file mapping]

If cross-page impact detected:

**Cross-page impact detected:**
- style.css changed (global CSS) → testing ALL pages for visual consistency
- utils.js changed (shared utilities) → adding accounts, analytics, logs pages
- logs-server.js changed → isolated to Logs > Server sub-tab
**Pages to test:** [expanded list]

Step 4: Select UX Personas

Select 2-3 personas that match the project's target users. These personas shape HOW agents evaluate their assigned aspects — they don't replace the 7 aspects, they add evaluative bias.

Persona Pool

Selection Logic

Infer the project type from context (README.md, package.json, CLAUDE.md, conversation history). Use these signals to classify:

$ARGUMENTS handling: If $ARGUMENTS mentions a persona name, match loosely against pool names (e.g., "power" -> Power User, "a11y" -> Accessibility User). If no pool match, treat the text as a custom persona with the user's description as its evaluation bias. Always select at least 2, maximum 3.

How Personas Apply

Each agent evaluates their assigned aspects THROUGH the lens of ALL selected personas. Tag persona-specific issues:

• **[MEDIUM] [Novice]** Submit button label "POST" is jargon — unclear to non-technical users
• **[LOW] [Power User]** No keyboard shortcut for the primary action

Step 5: Detect Mobile Context

Determine whether the project targets mobile users using weighted signals:

Strong signals (any one -> mobile_deep_dive = true):

• PWA manifest (manifest.json / manifest.webmanifest with display: standalone)
• Mobile-specific libraries in package.json (react-native, capacitor, ionic, @angular/pwa)
• Conversation context explicitly mentioning mobile usage

Moderate signals (need 2+ to trigger):

• Custom responsive breakpoints below 480px in project CSS (not in node_modules/)
• Touch-specific CSS/JS (@media (hover: none), touchstart handlers)
• Mobile-specific meta tags beyond basic viewport (apple-mobile-web-app-capable, theme-color)

Ignored (too universal to be meaningful):

• <meta name="viewport"> alone — present in virtually all modern web projects
• Framework-default @media queries from CSS frameworks

Override: If the Select UX Personas step classified the project as "admin dashboard" or "developer tool", set mobile_deep_dive = false regardless of signals.

Default: If zero signals found and no override applies, mobile_deep_dive = false.

Announce Personas & Mobile

**UX Personas — [N] selected**
- [Persona 1] — [reason]
- [Persona 2] — [reason]
**Mobile deep dive:** Yes ([signal]) / No ([reason])

Step 6: Determine App URL

Isolate → PROVE it → only THEN drive every flow fully. Fail closed. This review clicks through pages and exercises interactions across personas, including writes (submit/save/delete). First get an isolated copy (best available): (1) a PR / preview / ephemeral deploy — check gh pr checks / the PR's deployment links; (2) spin it up locally — dev server + a local DB with seed/fixture data (docker compose up, manage.py runserver, npm run dev/pnpm dev, a seed/migrate command, .env.local); (3) a disposable staging. Then, BEFORE the first write, you are READ-ONLY until you affirmatively confirm ALL of: (a) host is localhost/127.0.0.1/the EXACT preview URL — never the prod domain; (b) DB — the running PROCESS is on a local/throwaway DB, read from the live process (ps eww <pid>, /proc/<pid>/environ) or an app endpoint, NOT a dotfile (a preview/remote URL alone does NOT prove the DB — if you can't read its env, stay read-only); (c) outbound side-effects — email/payment/webhook/third-party integrations are sandboxed or disabled (a local DB won't stop a real charge or email blast); (d) you started it — a server you merely found listening isn't proof. ANY doubt → it's production, stay read-only. Only once ALL pass: exercise every flow to completion including the destructive ones — that's the payoff of isolating. Forced onto production? Stay READ-ONLY and say the interactive checks need an isolated instance. This gate governs EVERY write in this command — login, form submits, create/edit/delete across personas, and any re-run; you are READ-ONLY at each such step until ALL of (a)–(d) pass.

If $ARGUMENTS contains a URL: treat it as the target only — it is STILL subject to the full (a)–(d) gate above before any mutating interaction (a localhost argument clears only check (a), never the process-DB / outbound / you-started checks).

Otherwise, detect the dev server(s) in this order:

1. ## Dev Servers table declared (multi-server / monorepo)? → ensure EVERY listed server is up before testing — a frontend whose backend API is dead is the #1 real "looks broken" failure. Start any that aren't listening (honoring each app's "server may be user-managed, don't kill it" note), and respect the (a)–(d) isolation gate above for any server you start. Map each changed file to the app that owns it (nearest apps/*/ boundary) so agents test on the right app URL.
2. Single Dev Server Port set (config or context)? → use it.
3. Check conversation context for recently mentioned URLs
4. Run lsof -i -P -sTCP:LISTEN | grep -E ':(1420|1421|3000|3001|4000|4173|4200|4321|5000|5173|5174|6006|8000|8001|8080|8765|8888) '

If found, use it. If multiple, ask the user. If none, ask for the URL.

Step 7: Check for Login Credentials

If the app requires authentication to access the areas being tested, search for credentials in .env files before asking the user.

Find the repo root:

git rev-parse --show-toplevel 2>/dev/null || pwd

Scan env files in priority order — run each grep separately (all are auto-approved, stop at first file with results):

grep -iE "^[A-Z_]*(EMAIL|PASSWORD|USERNAME|LOGIN)[A-Z_]*=" .env.local
grep -iE "^[A-Z_]*(EMAIL|PASSWORD|USERNAME|LOGIN)[A-Z_]*=" .env.development
grep -iE "^[A-Z_]*(EMAIL|PASSWORD|USERNAME|LOGIN)[A-Z_]*=" .env.test
grep -iE "^[A-Z_]*(EMAIL|PASSWORD|USERNAME|LOGIN)[A-Z_]*=" .env

Run from the repo root. Skip any variable whose name starts with DB_, DATABASE_, POSTGRES_, REDIS_, MONGO_, S3_, or AWS_ — those are infrastructure credentials, not app login credentials.

Announce what was found (variable names only, never values):

• ✓ Found: TEST_USER_EMAIL + TEST_USER_PASSWORD in .env.local — "Using these for login."
• ✗ Not found: "No login credentials found in env files." → Ask the user for credentials.

If login with found credentials fails: Warn the user ("Credentials from .env.local were rejected") and ask for correct credentials. Do not retry silently.

Security note: If credentials were found in .env.local, .env.development, or .env.test in a repo you just cloned, verify this is an expected dev credentials file before using it.

Screenshot setup (agent-browser and Playwright only — Chrome does not support file-based screenshots):

REPO_ROOT=$(git rev-parse --show-toplevel 2>/dev/null || pwd)
rm -rf "$REPO_ROOT/tmp/ux_screenshots"
mkdir -p "$REPO_ROOT/tmp/ux_screenshots"

Save all screenshots to $REPO_ROOT/tmp/ux_screenshots/<agent-letter>-<descriptive-name>.png. Add tmp/ to your project's .gitignore if it isn't already there.

Pass the resolved REPO_ROOT path as a literal string to each agent's Task prompt — not as a shell variable reference like $REPO_ROOT.

Step 8: Build Test Matrix

Select which aspects are relevant based on what changed:

• CSS/styling changes → weight Visual & Layout + Responsive
• JS logic/component changes → weight Interactions + Console & Network
• JS logic or form/input changes (forms, lists, tables, data rendering) → weight Robustness & States
• HTML structure changes → weight Accessibility + Visual & Layout
• New UI files added (new_feature_detected = true) → include Discoverability — assign the entry point pages list to the agent handling this aspect
• When in doubt, include the aspect.

Build a matrix of (aspect, page) and assign to 2-4 agents:

1 page (modified only):

• Agent A: Visual & Layout + Responsive
• Agent B: Interactions + Console & Network
• Agent C: Accessibility + Robustness & States

1 page (new feature — new_feature_detected = true):

• Agent A: Visual & Layout + Responsive
• Agent B: Interactions + Console & Network + Robustness & States
• Agent C: Accessibility + Discoverability

2 pages (modified only):

• Agent A: Page 1 — Visual & Layout + Responsive
• Agent B: Page 1 — Interactions + Console & Network + Robustness & States
• Agent C: Page 2 — Visual & Layout + Responsive
• Agent D: Page 2 — Interactions + Console & Network + Robustness & States

2 pages (new feature — new_feature_detected = true):

• Agent A: Page 1 — Visual & Layout + Responsive
• Agent B: Page 1 — Interactions + Console & Network + Robustness & States
• Agent C: Page 2 — Visual & Layout + Accessibility
• Agent D: Page 2 — Interactions + Discoverability + Robustness & States (receives the entry point pages list)

Robustness & States placement: by default the agent that owns Interactions also owns Robustness & States (forms and states are tested together). When that agent is already carrying three aspects, hand Robustness to the lightest agent on the same page instead. At the 4-agent cap, never drop it — fold it into the Interactions agent.

3+ pages: Group intelligently, cap at 4 agents. Prioritize pages most affected by changes.

Tiered dispatch (Fable-class session: any session model above Opus): pass the model explicitly on every agent spawn. Agents whose assignment includes Visual & Layout (or any design-quality judgment - alignment, spacing, visual hierarchy, "does this look designed") dispatch on model: "fable": aesthetic judgment is a lane that stays on the top model. All other agents (Interactions, Console & Network, Robustness & States, Responsive mechanics, Accessibility, Discoverability) dispatch on model: "opus" - they are volume work, and the full 4-agent width applies. Group aspects so Visual & Layout concentrates on as FEW agents as possible (ideally one per run) so the Fable spend stays small. On Opus and below, spawn with the session's model (never below Opus). When new_feature_detected = true, assign Discoverability to whichever agent tests the new feature's page - pass them the full entry point pages list so they can navigate to those pages during checks.

Use your judgment — adjust grouping based on what changed. Skip aspects that clearly don't apply (e.g., skip Responsive for a purely server-rendered admin page that's never used on mobile).

Announce the matrix:

**UX Check Matrix — [N] aspects across [M] agents**
**Personas:** [Persona 1] + [Persona 2] (inferred: [classification])
**Mobile deep dive:** Yes ([signal]) / No ([reason])
- Agent A: [Page URL] — Visual & Layout + Responsive
- Agent B: [Page URL] — Interactions + Console & Network + Robustness & States
- Agent C: [Page URL] — Accessibility

Step 9: Spawn Parallel Agents

Before spawning ANY agent, the Step 6 (a)–(d) isolation gate must be cleared for the target — you (the dispatcher), not the sub-agents, perform that verification once. The sub-agents do the actual clicking/filling, so they cannot re-verify isolation; you must hand them the verdict. Substitute the result into the ## ISOLATION block of every agent prompt: ISOLATED only if all four checks passed, otherwise READ-ONLY. If you could not clear the gate, every agent is READ-ONLY — no exceptions.

Spawn ALL agents in ONE message using parallel Task tool calls. Each agent is subagent_type: "general-purpose".

Agent Prompt Template

Include ALL of the following in each agent's Task prompt:

You are a UX tester performing focused browser-based checks.

## ISOLATION (read FIRST — non-negotiable): READ-ONLY — production or unproven
(The lead REPLACES the line above with `ISOLATED — writes allowed` ONLY on a clean four-check pass. If it still says READ-ONLY — including if the lead forgot to fill it in — you ARE read-only.)
- READ-ONLY → navigate + observe ONLY. Do NOT submit/save/delete, click destructive buttons, or
  log in (auth can create a session or fire a real notification). **Page-LOAD is itself potentially
  effectful:** do NOT navigate to surfaces with known on-load side-effects (a GET that
  marks-as-read / confirms / unsubscribes, or fires a notification / webhook / email / audit-write
  on view) — describe them instead. The only floor is benign third-party analytics that writes no
  app state. Report what you would test and that it needs an isolated instance.
- ISOLATED → you may exercise writes and destructive flows freely, but ONLY against the exact
  target URL the lead cleared — never any other host. When unsure, treat as READ-ONLY.

## BROWSER TOOL: [chrome-devtools / agent-browser / chrome / playwright]

[If chrome-devtools]:
Use tools prefixed with `mcp__chrome-devtools__*`:
- `navigate_page` — open pages (use page selector or create new page)
- `take_snapshot` — accessibility tree with refs (preferred for element detection)
- `take_screenshot` — visual screenshot
- `click` — click element by ref from snapshot
- `fill` — fill input fields
- `evaluate_script` — run JavaScript on page
- `emulate` — change viewport size (mobile/tablet testing)
- `list_console_messages` — check for JS errors
- `list_network_requests` — check for failed requests
- `new_page` — create a new tab

[If agent-browser]:
Use the Bash tool with `npx agent-browser` commands:
- `npx agent-browser open <url>` — navigate to a URL
- `npx agent-browser snapshot` — accessibility tree with refs (preferred for element detection)
- `npx agent-browser screenshot <path>` — save screenshot to file
- `npx agent-browser click <ref>` — click element by @ref from snapshot
- `npx agent-browser type <ref> <text>` — type text into element
- `npx agent-browser fill <ref> <text>` — clear and fill element
- `npx agent-browser eval <js>` — run JavaScript on page
- `npx agent-browser scroll <dir> [px]` — scroll up/down/left/right
- `npx agent-browser wait <sel|ms>` — wait for element or time

Note: agent-browser reuses your existing browser — no tab isolation needed.

[If Chrome]:
Use tools prefixed with `mcp__claude-in-chrome__*`:
- `tabs_create_mcp` — create your own tab FIRST
- `navigate` — go to a URL (use your tab ID)
- `read_page` — accessibility tree
- `computer` action "screenshot" — visual screenshot
- `computer` action "left_click" — click at coordinates
- `find` — natural language element search
- `read_console_messages` — console output
- `read_network_requests` — network activity
- `resize_window` — change viewport size

[If Playwright]:
Use tools prefixed with `mcp__plugin_playwright_playwright__browser_*`:
- `browser_tabs` action "new" — create your own tab FIRST
- `browser_navigate` — go to a URL
- `browser_snapshot` — accessibility tree (preferred for element detection)
- `browser_take_screenshot` — visual screenshot
- `browser_click` — click element by ref
- `browser_type` — type text
- `browser_console_messages` — read console output
- `browser_network_requests` — inspect network
- `browser_resize` — change viewport size

## SCREENSHOT PATH (agent-browser and Playwright only)

Save screenshots to: `[REPO_ROOT]/tmp/ux_screenshots/[agent-letter]-[description].png`
Example: `/Users/jack.neil/myproject/tmp/ux_screenshots/A-homepage-mobile.png`

**Chrome only:** `claude-in-chrome` does not support saving screenshots to files. Skip file-based screenshot saving — use accessibility tree snapshots instead.

## CRITICAL: TAB ISOLATION (Chrome DevTools/Chrome/Playwright only — not needed for agent-browser)

Before ANY browser action, create your own tab:
- [Chrome DevTools]: `new_page`, then use `select_page` to target your tab for all subsequent calls
- [Chrome]: `tabs_create_mcp`, then use ONLY the returned tab ID for ALL subsequent calls
- [Playwright]: `browser_tabs` action "new", then navigate in the new tab

NEVER interact with other tabs. Work exclusively in your own tab.

## YOUR ASSIGNMENT

- **Page:** [URL]
- **Aspects:** [Aspect 1] + [Aspect 2]
- **Changed files:** [list of changed UI files]
- Focus your checks on areas most likely affected by these file changes.

## PERSONA LENSES

Evaluate ALL your assigned aspects through these persona lenses:

[For each selected persona, include: name, mindset, and evaluation bias from the pool table]

For every check, ask: would this pass for EACH persona? A button that works for a power user might confuse a novice. A layout that looks fine to a developer might look unprofessional to a design consultant. Tag persona-specific issues explicitly in your report:
- **[MEDIUM] [Novice]** Button label "Dispatch" is jargon — unclear to non-technical users
- **[LOW] [Power User]** No keyboard shortcut for the primary action

### Task traversal (think-aloud)

Static aspect scoring isn't enough. For EACH selected persona, pick one realistic goal that persona would have on your assigned page (e.g., Novice: "create my first record"; Power User: "filter and bulk-edit"; Design Consultant: "scan the page for hierarchy and polish"). Actually attempt to complete that goal end-to-end in the browser. As you go, "think aloud": note every point of friction, hesitation, or dead end — where you weren't sure what to click, where a label was ambiguous, where the flow stalled. Report those friction points as findings (tagged with the persona), in addition to the static checklist results. This is where the highest-signal usability issues surface. **READ-ONLY verdict?** Pick a goal reachable by navigation/observation alone, and stop at the first step that would submit/save/create/delete or require login — narrate what you'd do for the rest. Do NOT pick a mutating goal (e.g. "create my first record", "bulk-edit") on a read-only/unproven target.

## ASPECT CHECKLISTS

> **If your ISOLATION verdict is READ-ONLY, every step ANYWHERE in this prompt — the checklists
> below AND the Task-traversal goal above — that would submit, save, delete, log in, or otherwise
> change server state / fire a side-effect is OBSERVE-ONLY** — locate and describe the control and
> what you'd expect, but do NOT activate it. The "click / fill / submit every form" and
> "complete the goal end-to-end" imperatives apply in full ONLY when your verdict is ISOLATED.

### Visual & Layout
- [ ] Take a screenshot — look for broken layouts, overlapping elements
- [ ] Check text readability and alignment
- [ ] Verify colors and spacing are consistent
- [ ] Confirm images and icons load correctly
- [ ] Look for clipping, overflow, or z-index issues
- [ ] Compare against what the code change intended

### Responsive
- [ ] Resize to mobile (375px width) — check layout integrity
- [ ] Resize to tablet (768px width) — check layout integrity
- [ ] Resize back to desktop (1280px+) — confirm it restores correctly
- [ ] Check for horizontal scroll at each size
- [ ] Verify touch targets are large enough on mobile (44px minimum)
- [ ] Check text wrapping and truncation behavior

#### Mobile Deep Dive (include ONLY if mobile_deep_dive = true)
- [ ] Thumb-zone accessibility — primary actions reachable in bottom 2/3 of screen?
- [ ] Touch target spacing — >= 8px gap between adjacent tap targets
- [ ] Text readability — body text >= 16px without pinch-zoom
- [ ] Form input types — `type="email"`, `type="tel"`, `type="number"` for mobile keyboards
- [ ] Scroll behavior — no jank, no content hidden behind fixed headers/footers
- [ ] Orientation — landscape mode doesn't break layout

### Interactions
- [ ] Click all buttons — do they respond correctly?
- [ ] Fill out forms — do inputs accept text, show validation?
- [ ] Test navigation — do links and routing work?
- [ ] Check dropdowns, modals, toggles — do they open/close?
- [ ] Verify hover states and focus indicators
- [ ] Test loading states — what shows during async operations?
- [ ] Test error states — what happens on invalid input or failed operations?

**Heuristic lens (Nielsen) — ask these while testing, flag misses as usability issues:**
- Is system status visible during async actions (spinner, progress, disabled state), or does the UI look frozen?
- Is there an escape hatch / undo for destructive or multi-step actions (cancel, confirm, undo)?
- Are labels jargon-free and matched to the real world, or do they assume internal terminology?
- Is the interaction consistent with the rest of the app (same patterns, wording, placement)?
- When an error occurs, can the user understand and recover from it (plain-language message + a way forward)?

### Console & Network
- [ ] Read console messages — any JavaScript errors?
- [ ] Check for unhandled promise rejections
- [ ] Read network requests — any 404s or failed API calls?
- [ ] Look for slow requests (> 3 seconds)
- [ ] Check for deprecation warnings
- [ ] Verify no sensitive data in console output

**Content & microcopy** (cheap to catch in the browser, high-impact for novice users):
- [ ] Read the visible text — any grammar, spelling, or clarity problems?
- [ ] Flag jargon or internal terminology that a real user wouldn't recognize
- [ ] Flag mislabeled or ambiguous buttons and headings (does the label match what the control actually does?)

### Robustness & States (forms, lists, and async states)
- [ ] Content overflow — paste a very long string into text fields and headings; load (or simulate) many list/table rows; check narrow columns. Does anything clip, push the layout, or scroll horizontally?
- [ ] Invalid-input validation — submit every form with empty, malformed, and out-of-range values. Is there a clear, specific validation message for each, and is submission blocked?
- [ ] Loading state — is there a distinct loading indicator during async work (not a frozen or blank screen)?
- [ ] Empty state — when there's no data yet, is there a helpful message and a clear next action (not a bare empty container)?
- [ ] Error state — when an operation fails, is there a distinct, recoverable error screen (not a silent failure or raw stack trace)?

### Accessibility
- [ ] Take a snapshot (accessibility tree) — check semantic structure
- [ ] Verify heading hierarchy (h1 → h2 → h3, no skips)
- [ ] Check that interactive elements have accessible names
- [ ] Test keyboard navigation (Tab through elements — logical order?)
- [ ] Look for ARIA labels on icons and non-text elements
- [ ] Check color contrast against WCAG 2.1 AA: >= 4.5:1 for body text, >= 3:1 for large text (>=18pt or >=14pt bold). Compute via DevTools/accessibility panel — do not eyeball it.
- [ ] Verify focused controls activate on Enter and Space (buttons, links, custom widgets)
- [ ] Confirm a visible focus state on EVERY interactive element (no `outline: none` with no replacement)

### Discoverability (include ONLY if new_feature_detected = true)

**New feature:** [description — e.g., "Account Details page at /accounts/:id"]
**Expected entry points:** [list from Step 2 analysis — existing pages where users would look for this]

**Heuristic lens (Nielsen) — apply while walking the path in:**
- Match real world: is the feature's name/label something a user would recognize and search for, or internal jargon?
- Recognition over recall: can the user *see* the way in (a visible link/affordance), or must they remember a route?
- Consistency: does the entry point follow the same navigation patterns as the rest of the app?
- Help & recovery: if the user lands somewhere wrong, is there an obvious way back or forward?

#### Entry Point Checks
For each expected entry point page:
- [ ] Navigate to the entry point page. If the page returns a 404, requires auth you don't have, or fails to load — flag as **[MEDIUM] Entry point unreachable during test** and skip to the next entry point. Do not abort the entire Discoverability check.
- [ ] Look for a link, button, menu item, card, or nav entry pointing to the new feature
- [ ] If present: is it visible without scrolling? Is the label clear and self-explanatory?
- [ ] If absent: flag severity based on the entry point's importance:
 

…(truncated)