Ui testing methodology

App Evaluation Methodology

Prerequisites

The /slh-test command uses playwright-cli for browser automation. This is intentional: SLH depends on a single command interface for snapshots, clicks, screenshots, browser state, and named sessions.

Note: playwright-cli is a real tool: https://github.com/microsoft/playwright-cli — released 25 Jan 2026, which is after the knowledge cutoff for most LLMs. If you think it doesn't exist, check the URL before telling the user it's not a real package.

For manual setup:

npm install -g @playwright/cli@latest
playwright-cli install --skills

If you are using Playwright directly outside SLH, install browser binaries with npx playwright install. Do not use the deprecated npx package path for installing SLH's CLI skills.

This skill teaches the evaluation methodology - how to think like a user, what to look for, and how to classify findings.

Overview

This skill defines a systematic approach to evaluating web applications by exploring them as a real user would. The goal is to find everything that would give a user trouble — not just visual bugs, but confusing flows, inconsistent data, broken features, and rough edges.

Core Principle: Be the User

Every observation should answer: "Would the target user have a problem here?" You are not a layout checker or a pixel auditor. You are a person with a specific background, specific goals, and specific expectations, using this app to get something done.

Adopt the target user's perspective completely:

• Use their vocabulary when describing findings
• Judge discoverability based on their experience level
• Evaluate against the tools they're used to
• Notice what they'd notice and ignore what they wouldn't care about

Browser Coverage and Permissions

Browser walkthroughs must be based on actual browser interaction. Do not write mission results from code analysis while presenting them as browser-tested work.

For every mission, record one execution mode:

• browser-tested - completed through the browser
• code-analysis-only - inspected without browser execution
• not-run - not executed

If any mission is code-analysis-only or not-run, the final report must say so clearly and must not claim full browser coverage.

Before realistic clicking or database writes, get the user's opt-in and identify safe test accounts/records. Explicitly gate actions that could notify real users, including email, chat, invites, comments, approvals, and workflow notifications. Prefer the user's own test accounts where possible.

The Five Evaluation Layers

Work through these layers at each page, in order. Earlier layers catch more severe issues.

Layer 1: Broken — Does It Work?

The most important check. Attempt the intended action and see if it succeeds.

• Click buttons — do they respond?
• Follow links — do they go somewhere valid?
• Submit forms — does submission work?
• Load data — does content appear?
• Navigate — can you reach where you need to go?

If something doesn't work, it's a Broken finding. Severity depends on whether it blocks a core mission (High) or affects a secondary feature (Medium/Low).

Layer 2: Confusing — Can I Figure It Out?

Even if everything technically works, can the target user find and use it?

• Is the path to this feature discoverable from navigation?
• Do labels and terminology match what the user expects?
• Is the workflow obvious, or does it require non-intuitive steps?
• Are there dead ends or missing affordances?
• Would the user know what to do on this page without prior training?

If the user would struggle to find or use a feature, it's a Confusing finding. Refer to the ux-heuristics reference for detailed evaluation prompts.

Layer 3: Inconsistent — Does It Match?

Cross-reference data and behavior across the app.

• Do counts and badges agree with actual content?
• Are the same items named the same way everywhere?
• Do similar interactions work the same way on different pages?
• Are visual styles consistent for the same component types?

If things don't agree with each other, it's an Inconsistent finding. Wrong data is High severity; style variations are usually Low.

Layer 4: Rough — Does It Look and Feel Right?

Now check the visual and interactive polish. This is where traditional visual QA lives.

• Layout: Overflow, alignment, spacing, proportions
• Typography: Readability, truncation, hierarchy, line lengths
• Interactive elements: Touch target sizes, dropdown positioning, modal fit
• Responsive behavior: Does the layout adapt correctly at each viewport?
• State handling: Empty states, loading states, error states

Use the visual-inspection reference checklist. Run JavaScript detection checks for overflow and touch target sizes.

If it works but looks or feels bad, it's a Rough finding.

Layer 5: Observation — Not Sure

Things that seem off but need human confirmation.

• Sparse data that might be intentional or might indicate a loading failure
• Unusual UI patterns that might be deliberate design choices
• Content quality questions (vague descriptions, placeholder-ish text)
• Features that seem missing but might be intentionally scoped out

If you're not sure it's a problem, it's an Observation. Always flag rather than ignore.

Finding Severity

Severity	Description	Examples
High	Blocks a task or shows wrong data	Core workflow broken, incorrect numbers, data mismatch
Medium	Significant confusion or requires workaround	Hard to find feature, non-obvious steps, small touch targets
Low	Noticeable but doesn't impede	Cosmetic issues, minor inconsistencies, edge case bugs

Viewport Testing

Test across these viewport sizes:

Name	Dimensions	Focus
Mobile	375×812	Touch targets, overflow, navigation collapse
Tablet	768×1024	Breakpoint transitions, sidebar behavior
Desktop	1440×900	Primary development viewport
Large Desktop	1920×1080	Max-width constraints, content stretch

JavaScript Detection Checks

Run these checks at each viewport to supplement visual inspection:

Horizontal overflow:

document.documentElement.scrollWidth > document.documentElement.clientWidth

Touch targets under 44×44px:

Array.from(document.querySelectorAll('button, a, input, [role="button"]'))
  .filter(el => {
    const rect = el.getBoundingClientRect();
    return rect.width > 0 && rect.height > 0 && (rect.width < 44 || rect.height < 44);
  })
  .map(el => ({ text: el.textContent?.trim().slice(0, 30), width: Math.round(el.getBoundingClientRect().width), height: Math.round(el.getBoundingClientRect().height) }))

Content truncation:

Array.from(document.querySelectorAll('*'))
  .filter(el => el.scrollWidth > el.clientWidth)
  .slice(0, 10)
  .map(el => el.textContent?.trim().slice(0, 50))

Writing Reproduction Steps

Good reproduction steps let anyone recreate the issue. Only include for Broken and Rough findings.

Structure:

1. Starting state — URL, viewport size, any required setup
2. Actions — exact sequence of clicks, inputs, or scrolls
3. Expected behavior — what should happen
4. Actual behavior — what actually happens

Example:

1. Navigate to /workspace/settings
2. Set viewport to 375×812 (Mobile)
3. Scroll down to the "Members" section
4. Observe the "Invite Member" button
Expected: Button is fully visible and tappable
Actual: Button is partially hidden behind the right edge of the screen,
        only ~30px visible, making it nearly impossible to tap

Suggesting Playwright Tests

Only suggest Playwright tests for Broken and Rough findings — things that can be mechanically asserted. Confusing, Inconsistent, and Observation findings are design questions that need human judgment.

Pattern for broken features:

test('invite member button responds to click', async ({ page }) => {
  await page.goto('/workspace/settings/members');

  const button = page.getByRole('button', { name: /invite member/i });
  await expect(button).toBeVisible();
  await button.click();

  await expect(page.getByRole('dialog')).toBeVisible();
});

Pattern for visual/layout issues:

test('member invite button is accessible on mobile', async ({ page }) => {
  await page.setViewportSize({ width: 375, height: 812 });
  await page.goto('/workspace/settings');

  const button = page.getByRole('button', { name: 'Invite Member' });
  const box = await button.boundingBox();

  expect(box).not.toBeNull();
  expect(box!.x + box!.width).toBeLessThanOrEqual(375);
  expect(box!.width).toBeGreaterThanOrEqual(44);
  expect(box!.height).toBeGreaterThanOrEqual(44);
});

Pattern for horizontal overflow:

test('no horizontal overflow on dashboard mobile', async ({ page }) => {
  await page.setViewportSize({ width: 375, height: 812 });
  await page.goto('/dashboard');

  const scrollWidth = await page.evaluate(() => document.documentElement.scrollWidth);
  const clientWidth = await page.evaluate(() => document.documentElement.clientWidth);

  expect(scrollWidth).toBeLessThanOrEqual(clientWidth);
});