ui-test— agent skill

Preferences

_On startup, use Read to load ~/.claude/skills/ui-test/preferences.md. If missing, treat as first-run (see First-time detection)._

Defaults when no preferences exist:

• output-root: ./test-results/ui-test/ (cwd-relative when in a project; falls back to ~/.claude/skills/ui-test/runs/ outside a project)
• viewport: 1440x900 (single) — accepts a comma list for matrix runs (e.g. 375x812,1440x900)
• categories: e2e,visual,a11y,perf (debug is a separate mode, not a category)
• screenshot-strategy: per-step (per-step / on-fail / never)
• a11y-severity-floor: serious (minor / moderate / serious / critical — anything at-or-above is a failure)
• perf-budget: lcp<=2500,cls<=0.1,inp<=200,js-heap<=50mb — overridable per-run via --perf-budget
• verbosity: verbose-on-failure (concise / detailed / verbose-on-failure)
• spec-priority: inline-first (inline-first / project-first / both)
• auto-baseline: true (record on first run, diff thereafter)
• auto-open-report: true
• confirm-plan: always (always / only-multi-step / never)

_Also load ~/.claude/skills/ui-test/feedback-journal.md if present — surface the one-line Signal: from the latest 5 sessions to bias defaults (false-positive categories get pre-deselected, baselines flagged-as-flaky get re-recorded, etc.)._

Context

_On startup, use Bash to detect: today's date (date +%Y-%m-%d-%H%M), the current working directory (pwd), whether ./tests/ui/ exists, whether the user is inside a git repo (git rev-parse --is-inside-work-tree — silently fail if not), and whether the configured output-root exists. Do NOT call any browser tool yet — preflight runs in Step 1._

Command routing

Check $ARGUMENTS:

• `help` → display help, stop
• `config` → run config flow, stop
• `reset` → delete preferences.md + feedback-journal.md + sessions/ + resume-state.md, confirm, stop
• `feedback` → collect ratings on the most recent run (see Feedback section), stop
• `history` → list recent run folders with title, verdict, timestamp, stop
• `setup` → walk through Claude-in-Chrome extension setup (see Setup section), stop
• `record <name>` → take the most recent inline run and persist it as ./tests/ui/<name>.md, stop
• `--debug <url|description>` → enter interactive debug mode (see Debug section), stop
• `--file <path>` → load a single project test file and run it
• `--suite <glob>` → resolve the glob under ./tests/ui/, run each match in order
• anything else → treat $ARGUMENTS as an inline description, run it (Step 1 onward)

Help

ui-test — Drive real Chrome through plain-English UI tests

Usage:
  /ui-test "<description>"                Inline test, parsed and run
  /ui-test --file <path>                  Run a single ./tests/ui/*.md file
  /ui-test --suite "<glob>"               Run a glob of project tests
  /ui-test --debug "<url|description>"    Interactive debug session
  /ui-test record <name>                  Save the last inline run as a project file
  /ui-test history                        List recent runs (verdict + timestamp)
  /ui-test feedback                       Rate the last run (teaches the skill)
  /ui-test setup                          Walk through Chrome extension setup
  /ui-test config                         Set preferences
  /ui-test reset                          Clear preferences + journal + sessions
  /ui-test help                           This help

Per-run flags:
  --env <url> --viewport <WxH[,WxH]> --only <csv> --skip <csv>
  --baseline (force-record)  --perf-budget "lcp<=2000,cls<=0.05"
  --a11y-floor <minor|moderate|serious|critical>
  --headful-pause (pause for keypress before each navigation)
  --record-gif  --verbose
  --html-report (default on) | --no-html-report (skip the HTML report; print CLI only)
  --open-report | --no-open-report  (override auto-open-report preference)

Examples:
  /ui-test "log in as [email protected] on staging, expect Dashboard"
  /ui-test "homepage <2s load + passes a11y" --env https://acme.com --only perf,a11y
  /ui-test --suite "checkout/*.md" --viewport 375x812,1440x900
  /ui-test --debug "sidebar disappears on resize at /dashboard"
  /ui-test --file tests/ui/login.md --baseline

Output: {output-root}/{YYYY-MM-DD-HHMM}-{kebab-slug}/ contains
  report.html, data.json, spec.md, steps/, a11y/, perf/, visual/, sources.md.

Current preferences:
  (loaded from preferences.md)

Config

Use AskUserQuestion to collect, in batches of ≤4 per call:

Batch 1 — Output & viewport

1. Output root path (default ./test-results/ui-test/)
2. Default viewport(s) — comma-separated list
3. Auto-open report on completion (true/false)

Batch 2 — Categories & strictness

1. Default categories enabled (e2e / visual / a11y / perf — multiSelect)
2. Screenshot strategy (per-step / on-fail / never)
3. a11y severity floor (minor / moderate / serious / critical)
4. Perf budget (free text — lcp<=2500,cls<=0.1,inp<=200)

Batch 3 — Workflow

1. Spec priority (inline-first / project-first / both)
2. Confirm plan before run (always / only-multi-step / never)
3. Auto-record baseline on first visual run (true/false)

Save to ~/.claude/skills/ui-test/preferences.md:

# /ui-test preferences
Updated: {date}

## Defaults
- output-root: {path}
- viewport: {WxH[,WxH]}
- categories: {csv}
- screenshot-strategy: {per-step|on-fail|never}
- a11y-severity-floor: {minor|moderate|serious|critical}
- perf-budget: {lcp<=...,cls<=...,inp<=...}
- verbosity: {concise|detailed|verbose-on-failure}
- spec-priority: {inline-first|project-first|both}
- auto-baseline: {true|false}
- auto-open-report: {true|false}
- confirm-plan: {always|only-multi-step|never}

## Profile (optional — edit freely)
- Default base URLs by env: staging=, prod=, local=http://localhost:3000
- Standing exclusions (selectors / pages to ignore for visual diff):
- Standing a11y rule overrides (e.g., color-contrast: skip on /admin):
- Trusted slow third-parties (perf budget exclusions):

## Learned
<!-- Auto-appended from feedback over time -->

Confirm: "Saved. I'll use this as the baseline — and I'll keep sharpening as you rate runs."

Reset

Delete preferences.md, feedback-journal.md, sessions/, resume-state.md (each via Bash(test -f ... && rm) style — graceful on missing). Confirm: "All cleared. Starting fresh on the next run."

First-time detection

If no preferences file exists, show one warm non-blocking line:

First time running /ui-test — describe a test in plain English, I drive real Chrome via the Claude-in-Chrome extension and produce a per-run folder + single-file HTML report. Run /ui-test setup if the extension isn't connected yet, or /ui-test config to set defaults — otherwise continue and I'll auto-check the connection.

Then proceed to Step 1 (preflight). After the first successful run, inline-prompt for 1-2 quick prefs (categories, screenshot strategy) — don't force the full config flow.

Setup — Chrome extension onboarding

When invoked as /ui-test setup (or when Step 1 preflight fails and the user asks for help), load ~/.claude/skills/ui-test/reference/setup.md via Read and walk through it: prerequisites checklist (browser / plan / Claude Code version / extension version), enable steps (/chrome, /mcp), and a troubleshooting table. Reference: https://code.claude.com/docs/en/chrome — fetch via WebFetch if a detail in the local guide looks stale.

After the user says they're set up, run preflight (Step 1). On pass: "Connection good. Paste a test description: /ui-test \"<description>\"."

Workflow

Step 0 — Load learning context

Read ~/.claude/skills/ui-test/preferences.md and ~/.claude/skills/ui-test/feedback-journal.md. Continue silently if either is missing. Extract the latest 5 Signal: lines and apply them as biases (e.g., "user keeps flagging color-contrast as false positive on /admin → pre-deselect that rule for this run").

Step 1 — Preflight (Chrome extension check)

1. Call tabs_context_mcp.
2. Pass → continue silently to Step 2.
3. Fail (any error) → do NOT retry blindly. Stop and show:

Hmm — the Claude-in-Chrome extension isn't responding.

Common causes:
  • Extension not installed → https://chromewebstore.google.com/detail/claude/fcoeoabgfenejglbffodgkkbkcdhcgfn
  • Chrome integration not enabled → run  /chrome  → "Enable"
  • Extension idle → run  /chrome  → "Reconnect extension"
  • First time ever → restart Chrome once after enabling, then retry

Need the full walkthrough?  Run:  /ui-test setup

One retry allowed: if the user replies "retry", call tabs_context_mcp once more. If still failing, route to /ui-test setup.

Step 2 — Resolve the spec

Decide what to run from $ARGUMENTS, honoring spec-priority:

• Inline string → parse it as the spec (Step 3).
• `--file <path>` → load that file, parse its frontmatter (env, viewport, categories) + body.
• `--suite "<glob>"` → resolve under ./tests/ui/, queue each, run sequentially.
• No argument → if ./tests/ui/ exists and spec-priority is project-first or both, list project files via Glob and offer them via AskUserQuestion. Otherwise prompt: "Describe the test in one or two sentences."

Also extract per-run flags: --env, --viewport, --only, --skip, --baseline, --perf-budget, --a11y-floor, --headful-pause, --record-gif, --verbose.

Step 3 — Build the run plan

Convert the spec into a structured plan:

{
  title: kebab-slug,
  base_url: env || inferred,
  viewports: [{w, h}, ...],
  categories: filtered by --only/--skip and preference defaults,
  steps: [
    { id, kind: "navigate" | "click" | "fill" | "expect" | "wait" | "screenshot",
      target: selector | text | role,
      value: string | regex,
      timeout_ms: 5000,
      assertions: [{ type: "text" | "visible" | "url" | "console-clean" | "network-2xx", ... }] }
  ],
  perf_budget: parsed,
  a11y_floor: level,
  screenshot_strategy: from prefs unless overridden,
}

Write it to spec.md inside the run folder once Step 5 has created it. The plan is the source of truth — every later step references step IDs.

Step 4 — Confirm the plan

Honor confirm-plan preference:

• always → always show
• only-multi-step → skip if plan has 1 step + 0 assertions
• never → only echo a one-liner: "Running: {title} ({N} steps, {categories})"

Format:

ui-test plan — {title}
Base URL:   {url}
Viewport:   {WxH[, WxH...]}
Categories: e2e ✓  visual ✓  a11y ✓  perf ✓

Steps:
  1. navigate → /login
  2. fill #email = [email protected]
  3. fill #password = ••••
  4. click button[type=submit]
  5. expect h1 contains "Dashboard"
  6. screenshot full-page
  ...

Estimated: {N} browser ops, {M} screenshots, ~{seconds}s

  [Run]   [Tweak]   [Skip]

AskUserQuestion: "Run this plan?" — pre-select Run if Learned rules don't contradict.

Step 5 — Create the run folder

Slug: {YYYY-MM-DD-HHMM}-{kebab-title-max-60-chars}. Title summarizes the outcome being verified, not the URL: 2026-05-08-1430-login-flow-staging, 2026-05-08-1430-checkout-a11y-mobile.

mkdir -p {output-root}/{slug}/{steps,a11y,perf,visual}

Isolation rule: only the per-run folder is created/written. Never touch sibling folders. Baselines live at {output-root}/baselines/{title}/{viewport}/{step}.png so they're shared across runs of the same test.

Step 6 — Execute

For each viewport (matrix outer loop), for each step:

1. Tabs: call tabs_context_mcp once at the start of the run; create a fresh tab via tabs_create_mcp. Resize via resize_window to the current viewport. Never reuse tab IDs from a prior session.
2. Navigate: mcp__claude-in-chrome__navigate with the resolved URL.
3. Action:

• click / fill → prefer find to locate by role/text first, then javascript_tool for the actual interaction (more deterministic than relying on visual coordinates).
• fill on real form fields → form_input is fine for simple cases.
• wait → javascript_tool polling with timeout.

1. Assertion: javascript_tool returns a JSON { pass: bool, actual, expected, evidence }. Auto-assertions every step:

• read_console_messages since last step → fail on error level unless allowlisted in Profile.
• read_network_requests since last step → fail on 5xx, optionally 4xx based on prefs.

1. Screenshot (per screenshot-strategy): use javascript_tool to call html2canvas (load via CDN if missing, single attempt; fall back to logging the gap) and write the PNG to steps/{NN}-{stepId}.png. If a step fails, screenshot regardless of strategy.
2. Per-step JSON: append to data.json — { id, status, duration_ms, screenshot, console, network, assertions }.
3. Stop conditions: a hard-fail step (e.g., navigation failure) halts subsequent steps for that viewport and marks them skipped. Other viewports continue.

Browser dialog rule (from MCP server instructions): if any step would trigger an alert/confirm/prompt, refuse to run it and mark the step blocked with a clear note. Never click elements that may surface a modal — the extension becomes unresponsive.

Batching: when many similar reads are needed in a row (selectors, text), prefer browser_batch over a serial loop.

Optional GIF: if --record-gif, wrap the run with gif_creator start/stop calls and save to run.gif.

Step 7 — Per-category extras

After the main step loop, run category-specific passes:

• a11y — inject axe-core via javascript_tool:

  const s = document.createElement('script');
  s.src = 'https://cdnjs.cloudflare.com/ajax/libs/axe-core/4.10.0/axe.min.js';
  document.head.appendChild(s);

Wait for window.axe, run axe.run(), write a11y/violations.json. Filter by a11y-severity-floor. If CDN load fails, log the gap and skip — don't fake-pass.

• perf — javascript_tool reading performance.getEntriesByType('navigation'), PerformanceObserver for LCP/CLS/INP, performance.memory for JS heap. Combine with read_network_requests for waterfall. Write perf/metrics.json. Compare against perf_budget; mark each metric pass/fail.

• visual — for each screenshot step: if auto-baseline=true and no baseline exists, save as baseline; else compute pixel diff (use javascript_tool with a tiny inline diff implementation, or skip diff and just save the new snapshot if diff lib unavailable). Write visual/{step}.png + visual/{step}.diff.png. Threshold from prefs (default 0.1% pixel delta).

Each category degrades gracefully — a CDN block, a browser quirk, or a missing API does NOT fail the run. It logs the gap in sources.md and shows up in the report's "Coverage gaps" panel so the verdict is honest about what was actually checked.

Step 8 — Verdict and report

Compose verdict per run:

• All steps pass + all categories pass → PASS — {title}
• Any e2e step fail → FAIL — {first-failing-step-summary}
• Only category fails (a11y/visual/perf) → PASS WITH ISSUES — {N} categories below bar
• Coverage gaps prevented a verdict → INCONCLUSIVE — {gap reason}

Generate report.html (skip if --no-html-report). Single self-contained file: inline JS ≤200 lines, no external scripts, CSS variables only, dark + light via prefers-color-scheme, mobile-responsive, bento-grid layout (12-col grid with mixed-size cards) reflecting current 2026 dashboard trends. Verdict-forward: hero card with verdict pill + screenshot, perf bento row (4 metric cards with conic-gradient rings + sparklines), step timeline with filter chips, a11y panel with severity badges, visual diff strip, coverage panel, footer.

Skeleton source — copy the structure and styles from ~/.claude/skills/ui-test/examples/sample-report.html (a fully-rendered reference with mock data). Replace its hardcoded data with the run's actual values. Load ~/.claude/skills/ui-test/reference/report.md for the full layout spec and inline-JS responsibilities if more detail is needed.

Step 9 — CLI summary

Print to terminal (verbosity-aware):

ui-test — {title}

  {VERDICT}   {one-line summary}

  Steps:    {pass}/{total} passed  ({fail} failed, {skip} skipped)
  a11y:     {N} violations ≥ {floor}        ({pass|fail})
  perf:     LCP {ms}ms · CLS {n} · INP {ms}ms   ({pass|fail})
  visual:   {changed}/{total} snapshots over threshold

Folder:   {output-root}/{slug}/
Report:   {output-root}/{slug}/report.html

  [Open report]    [/ui-test feedback]    [/ui-test record <name>]

If auto-open-report=true, run open {report.html}.

Step 10 — Save session log + invite feedback

Append a one-page session log to ~/.claude/skills/ui-test/sessions/{slug}.md:

# {slug}
Date: {timestamp}
Spec: {one-line summary or path}
Verdict: {label}
Failures: {short list}
Categories run: {csv}
Coverage gaps: {csv if any}

End with one soft line:

Run /ui-test feedback — even one rating sharpens future runs (esp. flagging false-positive a11y rules or flaky baselines).

Feedback & learning

When invoked as /ui-test feedback:

1. Find the most recent file in sessions/. If none, say No recent run found. and stop.
2. Print a one-line summary of that session.
3. Ask via AskUserQuestion (3-4 questions, one batch — pre-select using Learned):

• Was the verdict right? (correct / false-positive / false-negative / inconclusive — needed more)
• Which category misfired (if any)? (none / e2e / visual / a11y / perf — multiSelect)
• Screenshot strategy felt: (just right / too many / too few)
• Anything to flag for next time? (free text — flaky selector, baseline regen, perf-budget tweak, env-specific override)

1. Append to ~/.claude/skills/ui-test/feedback-journal.md:

## {slug} — {date}
- Verdict: {label}
- Correctness: {answer}
- Misfiring categories: {csv}
- Screenshot strategy: {answer}
- Notes: {free text}
- Signal: {one-line generalization, e.g., "color-contrast on /admin keeps false-positive — exempt rule"}

1. Promotion rule — if 3+ sessions show the same Signal: (substring match on the key phrase), promote it to ## Learned in preferences.md and mention once: "Noticed {pattern} across recent runs — saving as a standing default."
2. Drift correction — if 2 newer sessions contradict a Learned rule, demote it (move to journal with a Demoted: <reason> line). Never leave stale rules in Learned.

The journal is human-editable — respect anything the user writes there directly.

Debug mode (--debug)

Invoked as /ui-test --debug "<url-or-description>". Skips the assertion-driven flow; instead, runs an investigative session:

1. Preflight (same as Step 1). If the input is a description without URL, infer or ask once for the URL.
2. Open a fresh tab at the URL. Resize to the configured viewport.
3. Capture a baseline of: full-page screenshot, console (last 100 messages), network (all requests since navigation), DOM stats (node count, deep nesting, common a11y issues spot-check).
4. Loop: ask the user one AskUserQuestion at a time — [Reproduce now] (waits for the user to act in the browser, then re-snapshots), [Resize {viewport}], [Run a11y], [Run perf], [Inspect selector], [Tail console for 5s], [Done — write report].
5. Each iteration writes deltas to debug/iter-NN.{png,console.json,network.json}.
6. On [Done], generate report.html (same template, "Debug" verdict variant — focuses on findings, not pass/fail) and a findings.md with prioritized issues (Critical / Medium / Low).
7. End with: "Want to convert this into a regression test? Run /ui-test record <name>."

Debug mode never asserts and never fails — it surfaces.

Record mode (record <name>)

Invoked as /ui-test record <name>:

1. Locate the most recent inline run (or debug session). If none, say so and stop.
2. Generate ./tests/ui/<name>.md (creating the directory if missing) with:

---
title: <Human Title>
env: {base url}
viewport: 1440x900
categories: e2e,a11y
---

# <Human Title>

> Recorded from run {slug} on {date}.

## Steps

1. Navigate to /login
2. Fill #email = [email protected]
3. ...

## Assertions

- Page contains heading "Dashboard"
- No console errors
- All network requests 2xx-3xx

1. Confirm path. Suggest: "Run with /ui-test --file tests/ui/<name>.md or include in a suite."

Filename / folder rules

• Run folder: {YYYY-MM-DD-HHMM}-{kebab-title} — title ≤ 60 chars, summarizes the verification, not just the URL
• Step file: {NN}-{kebab-step}.png — NN zero-padded order
• Baseline: {output-root}/baselines/{title}/{viewport}/{step}.png — shared across runs of the same test
• Project test: ./tests/ui/<kebab-name>.md — never .test.md suffix
• Never append -2, -3 for collisions; differentiate by viewport or env

Principles

1. Plain English in, structured plan out — every run materializes a spec.md so the user can see exactly what was parsed before anything moves.
2. Real Chrome, explicit tabs — always tabs_context_mcp first; always create a new tab; never reuse a tab ID from a prior session.
3. Honest about gaps — if axe-core's CDN was blocked or a screenshot library failed, the report says so. No fake passes.
4. Verdict-forward report — the HTML hero is the verdict, not the data grid.
5. Per-run isolation — only the run folder is mutated; baselines live in their own shared folder. Never delete sibling runs.
6. Stop on dialogs — refuse to interact with elements that surface alert/confirm/prompt; mark the step blocked rather than freezing the extension.
7. Graceful degradation — missing preferences, missing journal, blocked CDN, missing baseline all degrade with a logged gap, not a failure.
8. Destructive actions need confirmation — reset and --baseline (overwrites baselines) print what they'll touch and require a yes.
9. Learn quietly — promote a rule to Learned only after 3+ matching signals; demote on 2 contradictions; mention promotions and demotions once.
10. Stop means stop — when the user halts mid-run, save resume-state.md with the next step ID and invite /ui-test resume (future enhancement; not yet a subcommand).
11. Skill orchestrates, doesn't re-judge — preserve the user's spec verbatim in spec.md; never silently rewrite assertions.