exploratory-qa

name

exploratory-qa

description

Senior-engineer exploratory QA for full-stack apps built with the design-to-nextjs + nestjs-enterprise-backend stack. Drives Playwright through the real app to find what automated tests miss — missing empty states, oversized buttons inconsistent across pages, overlapping z-indexes, frozen frames during large data loads, client-side filter/sort that should be server-side, three modules that should share one table component, slow queries with N+1, missing loading states, error states never tested, mobile breakpoints that explode, focus traps in dialogs, cross-module refactor candidates. Six-phase pipeline (environment, happy-path, edge-cases, consistency, performance, report) with four subagents (qa-environment, qa-flow-tester, qa-consistency-checker, qa-performance-prober). Produces QA_REPORT.md with severity-ranked findings, evidence, refactor recommendations. Use for production-readiness QA, UX audit, design consistency, load testing, exploratory testing.

Exploratory QA

Senior-engineer-style exploratory QA: drive the real app like a discerning user, notice everything wrong, and produce a report a team can actually act on. This is not unit-test generation — it's "I spent two days clicking through your app and here's what's broken, inconsistent, or going to embarrass you in front of real users."

When to use this skill

Use when:

• The build is functionally complete but you need a production-readiness check
• A user / customer / investor demo is coming up and you want to find issues before they do
• You suspect there are inconsistencies across pages but don't know where
• You're worried about performance with real data volumes
• The app was built by AI/agents and you want a human-style pass over the output
• The team needs an honest report of what's incomplete, missing, or visually wrong

Do NOT use this skill for:

• Unit test generation — Jest/Vitest skills handle that
• Functional regression suites — that's integration-tester from the orchestrator
• Security audit — that's security-review-and-fix
• Initial build — use the orchestrator or prototype-to-saas

This skill assumes a working app exists and runs. It explores; it doesn't build.

How to invoke this skill

This skill is designed for Claude Code. It installs 4 specialized subagent definitions into .claude/agents/ and the main session orchestrates them through six phases. The flow-tester agent drives Playwright; have Playwright available (pnpm dlx playwright install chromium) before kicking off.

Pattern 1 — natural language (standalone QA pass)

Start a Claude Code session in the project's directory:

Run a production-readiness QA pass on this app. The stack is up at
localhost:3000 (web) and localhost:3001 (api).

The main session reads this skill's SKILL.md, installs the 4 QA subagents via the install script, and runs the six phases by delegating to subagents through natural language.

Pattern 2 — invoked from the orchestrator

When prd-design-build-orchestrator runs end-to-end and this skill is installed, the QA pass automatically runs as Step D.3 of Phase D. No separate invocation needed.

Pattern 3 — agent-teams mode for specific phases

This skill defaults to subagents. Two phases benefit from agent teams:

• Phase 6 (report synthesis) — 3-teammate adversarial team that debates each finding's severity before writing QA_REPORT.md
• Performance investigation for hard issues — competing-hypotheses team when a perf issue has multiple plausible causes

Both require CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. See "Agent teams (optional)" at the end of this skill for the invocation prompts.

The 4 QA agent definitions work as both subagent definitions AND teammate types without modification.

What this skill catches that automated tests don't

Things automated tests usually pass but humans notice immediately:

• Missing empty states — table just shows headers and nothing else when zero rows
• Inconsistent primitive sizing — <Button> on /companies is lg, on /contacts is default, on /campaigns is sm — same semantic action, three sizes
• Z-index collisions — toast shows behind dialog backdrop, dropdown hidden under sticky header
• Frozen UI during data load — list page imports 5000 rows of fixtures, filter input is unresponsive while rendering
• Client-side data work that should be server-side — .filter().sort() over thousands of records in useMemo blocking the main thread
• Refactor opportunities — three modules each have a near-identical <DataTable> with slight differences; should be one shared component
• Missing loading skeletons — page flashes empty before content arrives
• Missing error states — API returns 500, page just stays in loading state forever
• Overflow disasters — long company name breaks the row layout, long text spills past container
• Focus traps broken — <Dialog> doesn't trap focus, Tab escapes to background
• Mobile breakpoints exploding — sidebar overlays content; modal extends past viewport
• Inconsistent headers — every page renders its title with different typography/spacing
• Async race conditions — fast-clicking the create button submits twice
• N+1 query problems — list of 50 companies fires 51 SQL queries server-side
• Stale data after mutation — created a company, list doesn't refresh, user thinks it failed
• Form validation timing — errors flash on every keystroke instead of on blur

These are the things that distinguish "I shipped" from "I shipped something I'd be proud to show."

The six-phase pipeline

Phase 1: ENVIRONMENT      Verify stack, plant test data at realistic scale, capture baselines
Phase 2: HAPPY PATH       Drive canonical flow for every feature; record observations + screenshots
Phase 3: EDGE CASES       Empty/loading/error states; large data; slow network; failed mutations
Phase 4: CONSISTENCY      Cross-route checks: button sizing, spacing, headers, z-index, refactor smells
Phase 5: PERFORMANCE      Load times, layout shift, frozen frames, database query counts
Phase 6: REPORT           QA_REPORT.md with severity-ranked findings + recommendations

Phases 2, 3, 4 are largely parallel-safe — they share no inputs and write to different scratch directories. Phase 5 needs exclusive control of the running stack (performance measurement is sensitive to other activity). Phase 6 is the orchestrator's synthesis.

The four agents

Agent	Phase	Tools	Owns
qa-environment	1	Read, Glob, Bash	seed data, baseline state, QA_ENVIRONMENT.md
qa-flow-tester	2 + 3	Read, Bash, Glob, Grep	qa/flows/<feature>/ directory with screenshots, traces, observations
qa-consistency-checker	4	Read, Glob, Grep, Bash	qa/consistency/ with cross-route diff data + QA_CONSISTENCY.md
qa-performance-prober	5	Read, Bash	qa/performance/ with timing data + QA_PERFORMANCE.md

Plus reuse of integration-tester from the orchestrator skill for the pre-flight functional smoke (if integration tests fail, exploration is pointless — fix functional issues first).

Five agents would feel like the right number to match other skills, but a fifth agent here would just split work artificially. Four works.

All agent definitions in references/qa-agents.md. The standalone installer at references/install-qa-agents.sh extracts them via the orchestrator's extract-agent.py.

Phase 1 — Environment

Goal: the stack is up, the data is realistic, the baselines are captured.

Install agents:

# Reuses the orchestrator's extract-agent.py
~/.claude/skills/exploratory-qa/references/install-qa-agents.sh

Then dispatch in natural language:

"Use the qa-environment subagent to verify Docker + API + worker + web are running and healthy. Seed test data at REALISTIC SCALE (not toy size — see the seed-scale rules in ~/.claude/skills/exploratory-qa/references/environment-checklist.md). Capture baselines: take screenshots of every route at 1440px, 768px, and 375px viewports. Plant the edge-case fixtures listed in the checklist. Produce QA_ENVIRONMENT.md."

The "realistic scale" matters. A test DB with 5 companies hides the performance and overflow issues that show up with 500. The agent seeds to numbers like:

• Workspaces: 3 (test cross-workspace isolation alongside the audit)
• Users per workspace: 5
• Companies: 250 per workspace (realistic SMB customer)
• Contacts: 8 per company → ~2000 per workspace
• Campaigns: 30 per workspace (mix of statuses)
• Leads + deals: enough to populate a multi-column kanban with realistic density
• Email sent: 50k+ rows per workspace (exercises the hypertable)

The exact numbers come from MIGRATION_PLAN.md if it exists, or EXECUTION_PLAN.md if greenfield, or from sensible defaults in references/environment-checklist.md.

Baseline screenshots become the reference for Phase 4's consistency checks.

Phase 2 — Happy Path

Goal: the canonical flow works end-to-end for every feature. Find what's broken even on the easy case.

The main session dispatches one qa-flow-tester per feature, in parallel batches of up to 6. For a 6-feature wave:

"Dispatch six qa-flow-tester subagents in parallel, each scoped to one feature module (companies, contacts, mailboxes, campaigns, pipeline, inbox). Mode: happy-path. Each should follow the recipes in ~/.claude/skills/exploratory-qa/references/flow-recipes.md — login → list page → filter → sort → detail → create → toast verification → list refresh. Take screenshots at every step. Write to qa/flows/<feature>/happy-path/."

For projects with more than 6 features, run in successive batches; do not exceed 6 concurrent subagents per tool-use turn.

Each flow tester writes to its own directory:

qa/flows/companies/
  ├── screenshots/
  │   ├── 01-list.png
  │   ├── 02-list-filtered.png
  │   ├── 03-detail.png
  │   ├── 04-add-dialog.png
  │   └── 05-list-after-create.png
  ├── trace.zip                  (Playwright trace)
  ├── network.har                (network log)
  ├── console.log                (browser console output)
  └── observations.md            (what the agent noticed)

observations.md is the structured output: timed steps, visual issues spotted, console errors, network anomalies. The orchestrator collates these in Phase 6.

Phase 3 — Edge Cases

Same qa-flow-tester agent, different prompts. Now the agent deliberately tries to break things.

Categories per feature:

• Empty data — clear all rows, verify empty state renders correctly (not just a blank page)
• Single item — verify list with one row, detail with no related items
• Large data — load with 1000+ items, check render time, scroll performance, filter responsiveness
• Slow network — Chrome devtools throttling to "Slow 3G," verify loading skeletons appear
• API errors — mock the API to return 500, verify error state renders (not just stuck on loading)
• Validation errors — submit form with missing required fields, verify inline errors, check that errors don't flash on every keystroke
• Concurrent mutations — fast-click submit button, verify only one POST fires (debounce or disable-while-pending)
• Stale data — open two tabs, mutate in tab A, refresh tab B, verify update visible
• Long content — extremely long names, descriptions, URLs — verify no overflow disasters
• Special characters — names with emoji, Unicode, HTML entities — verify rendering and persistence
• Boundary numbers — quantity = 0, very large numbers, negative numbers in inputs that should reject them
• Browser back/forward — navigate via browser controls, verify state is preserved
• Tab/keyboard nav — Tab through a form; verify logical order; Escape closes dialogs; focus traps work in modals
• Mobile breakpoints — same flows at 375px width; check for overflow, layout breakage, untouchable buttons

Each category becomes its own observation file under qa/flows/<feature>/edge-cases/. The agent flags severity (Critical/High/Medium/Low) inline.

Phase 4 — Consistency

Cross-cutting analysis across the whole app. The qa-consistency-checker agent reads the source + the Phase 1 baseline screenshots + the Phase 2/3 observations.

Visual consistency

For every component used in 2+ places, compare:

• Computed font sizes / weights / line-heights
• Computed padding / margin around the primitive
• Computed border-radius / shadow
• Hover / focus state appearance

Example findings:

• "Primary <Button> is size="lg" (height 44px) on /companies/new but size="default" (height 40px) on /contacts/new. Inconsistent."
• "Page title <h1> is text-2xl font-semibold on /companies but text-3xl font-bold on /contacts. Should be one component."
• "Table row hover background is bg-muted on /companies but bg-accent on /contacts."

Structural consistency

Source-code analysis for refactor opportunities:

• Three modules each define their own <DataTable>, <PageHeader>, <StatCard>, <EmptyState> — should be shared in apps/web/src/components/shared/
• The same data-fetching pattern (useQuery + filter state in URL) is re-implemented per module — extract a useListQuery hook
• Multiple confirm-delete dialogs exist with slightly different wording — should be one <ConfirmDeleteDialog> component

Layout consistency

• Every <aside> sidebar should be the same width (we now use shadcn's block, but the QA verifies in practice)
• Every page in the (authed) layout has the same content padding
• Modal/dialog widths are from a consistent scale (sm/md/lg/xl), not ad-hoc px values
• Z-indexes used: should be a finite ordered scale, not random z-50 here and z-100 there

Behavioral consistency

• Every list page implements filter+sort+paginate the same way
• Every mutation surfaces success the same way (sonner toast position + duration)
• Every form's validation timing matches (on blur, or on submit, not both inconsistently)
• Every keyboard shortcut works across pages (Esc closes dialogs everywhere, not just some)

The agent produces QA_CONSISTENCY.md with findings grouped by category, each citing exact file:line for source issues and screenshot diffs for visual ones.

Phase 5 — Performance

The qa-performance-prober agent runs instrumented flows and measures.

Per-page metrics

For each route:

• Time to first content — <main> becomes non-empty
• Time to interactive — buttons respond to clicks
• Largest Contentful Paint — the standard Web Vitals metric
• Cumulative Layout Shift — does the page jump after first render?
• Total Blocking Time — main thread blocked time during page load
• JS bundle size — Next.js build output for this route

Per-flow timing

For each happy-path flow from Phase 2:

• Total flow duration
• Time per step
• Network requests fired (count + total bytes)
• Frames dropped during interactions (scroll, drag, animate)

Database query counts

For each frontend-initiated action, count backend queries:

-- Enable on dev DB
SET log_statement = 'all';
-- Run the flow
-- Tail the Postgres log; count queries per HTTP request

Flag:

• N+1 patterns — list endpoint fires N queries where 1 join would do
• Missing eager loads — detail endpoint hits the DB 8 times instead of 2
• Full-table scans — query plan shows Seq Scan on a table with an indexable column
• Unbounded results — query without LIMIT running against a large table

Frontend data-volume analysis

The reason your useMemo example matters:

• Component renders <Table> from useQuery data
• Data array length is N
• Filter/sort applied via .filter().sort() in useMemo
• Test: at N=5000, type a character in the filter input — frame budget?

Flag any pattern where:

• Data length is dynamic (not bounded to <100 by pagination)
• Filter/sort is client-side
• The filter input doesn't debounce

These cause UI freezes that don't show up in functional tests but feel terrible to users.

Memory / leak detection

Long-running interactions (5 minutes of clicking around) — does heap usage grow unbounded? Likely places: chart libraries forgetting to dispose, WebSocket subscriptions not cleaning up, image listeners.

The agent produces QA_PERFORMANCE.md with timing tables, query counts, flagged anti-patterns.

Phase 6 — Report

The orchestrator (not a subagent) reads everything:

• qa/flows/*/observations.md (every happy + edge case observation)
• qa/consistency/QA_CONSISTENCY.md
• qa/performance/QA_PERFORMANCE.md
• Screenshots, traces, network logs for evidence references

And writes QA_REPORT.md per the template in references/qa-report-format.md.

The report:

• Executive summary (counts by severity)
• Per-feature breakdown (what's solid, what's broken)
• Cross-cutting findings (consistency, performance, refactors)
• Recommended ordered fix list
• Evidence appendix (screenshots, traces referenced by ID)

Severity ladder:

Severity	Definition	Examples
Critical	Blocks launch — feature is broken or major UX failure	Form doesn't submit; create button shows wrong dialog; mobile sidebar covers content uncloseable
High	Embarrassing in front of users; noticeable in normal use	Missing empty state; 4-second page load; toast hidden behind dialog; broken on iPhone
Medium	Inconsistency or polish issue that compounds	Button sizes vary across pages; page titles inconsistent; minor layout shifts
Low	Nice-to-fix; doesn't impact real users	Hover state subtle inconsistency; 200ms delay where 100ms would feel snappier
Refactor	Not a bug; a code-quality opportunity	Three modules duplicate DataTable; useListQuery hook waiting to be extracted

Critical findings should block launch. High findings need explicit user acknowledgment. Medium/Low/Refactor go in the report as recommendations.

Integration with the orchestrator

When the user runs the build orchestrator end-to-end, this skill slots into Phase D (after the security review):

Phase D — INTEGRATE (orchestrator flow):
  D.1  integration-tester       (functional smoke)
  D.2  security-review-and-fix  (security audit, if installed)
  D.3  exploratory-qa           (THIS skill, if installed)
  D.4  final report (orchestrator synthesizes everything)

A failed integration test stops the pipeline before security/QA runs (no point exploring broken). Unresolved Critical findings in the QA report block the final orchestrator report.

If this skill is run standalone (post-launch quality pass), it's invoked directly without the orchestrator.

Reference files

File	When to read
references/qa-agents.md	Phase 0 install — source-of-truth for all 4 agent definitions
references/environment-checklist.md	Phase 1 — seed scale + baseline capture procedure
references/flow-recipes.md	Phase 2 + 3 — Playwright snippets for common flows (login, list, create, update, delete, etc.) and edge-case scenarios
references/consistency-checklist.md	Phase 4 — what to compare across pages, what counts as a finding
references/performance-checklist.md	Phase 5 — metrics to capture, anti-patterns to flag, query-count methodology
references/qa-report-format.md	Phase 6 — QA_REPORT.md template
references/install-qa-agents.sh	Phase 0 install script

Validation checklist

Before declaring QA complete:

• .claude/agents/ contains the 4 QA agents
• QA_ENVIRONMENT.md exists with seed counts and baseline screenshots
• qa/flows/<feature>/ exists for every feature module (happy + edge)
• QA_CONSISTENCY.md exists
• QA_PERFORMANCE.md exists with timing tables
• QA_REPORT.md exists with severity-ranked findings
• All Critical findings either resolved or explicitly accepted by user
• Screenshots are referenced by file path in the report (reviewer can verify)
• Database query counts are recorded for at least one full happy-path flow
• Mobile breakpoint coverage: 375px tested for every list/detail/form
• Performance baselines recorded (LCP, CLS, TBT per route)

Common pitfalls

P1 — Running with toy data. A 5-row test DB hides 80% of what this skill exists to find. Phase 1's seed scale matters; don't shortcut it.

P2 — Skipping mobile. Most issues live at 375px width. The flow-tester should re-run every flow at mobile viewport, not just desktop.

P3 — Treating "no console error" as "no issue." Many UX issues produce zero errors. The agent must observe behavior + visual state, not just console output.

P4 — Reporting noise. A 200ms render diff on a primitive isn't a finding; a 2-second freeze on filter is. The agent applies severity rubric, not just "everything different."

P5 — Trying to test things automated tests already cover. Don't re-run unit tests. Don't re-run integration tests. Focus on what only an exploring engineer would find.

P6 — Forgetting to flag refactor opportunities. This skill is a unique vantage point for cross-module patterns. Three modules with duplicated table code = one refactor finding, not three "bug" findings.

P7 — Reporting without evidence. Every finding cites a screenshot path, source file:line, trace ID, or query log excerpt. "Companies page is slow" is a useless finding; "Companies list with 250 rows: LCP 3.8s, TBT 1.1s, see qa/performance/companies-list.trace.zip" is actionable.

P8 — Running Phase 5 alongside Phase 2/3. Performance measurement needs an idle system. Run perf last, after all flow testing is done and the system has settled.

Agent teams (optional, experimental)

Most of this skill's pipeline fits the subagent pattern cleanly — per-feature flow-testing parallelizes, then a single consistency-checker and performance-prober follow. Two phases benefit from agent teams when stakes are high.

Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in settings.json or environment.

Adversarial report synthesis (Phase 6 alternative)

The default Phase 6 has the orchestrator read all observations and write QA_REPORT.md with severity assignments. A single perspective can be wrong about what blocks launch versus what's polish.

After Phases 1-5 finish, instead of synthesizing as orchestrator, ask the main session:

"Spawn a 3-teammate agent team for the QA report synthesis: a harshest-critic teammate that pushes every Medium up to High by default, a pragmatist teammate that pushes back on inflated severities, and a shipping-advocate teammate that argues what genuinely blocks launch versus what's polish. Have them read all qa/flows/*/observations.md, QA_CONSISTENCY.md, and QA_PERFORMANCE.md, and debate every finding's severity before recording it. Produce QA_REPORT.md by consensus, with each Critical and High finding requiring sign-off from all three teammates."

Useful when the report's recommendations drive launch/no-launch decisions or budget for a fix sprint.

Competing-hypotheses for hard performance issues

When qa-performance-prober flags an issue with multiple plausible causes (the canonical example: "companies list LCP 3.8s with 250 records — could be N+1 query, large bundle, unindexed column, or client-side filter blocking the main thread"), a single investigator picks one theory and may stop there.

On-demand for a specific issue:

"Spawn a 3-teammate agent team to investigate the companies list LCP issue: one teammate investigating the N+1 hypothesis (query logs, Postgres EXPLAIN), one investigating bundle size (Next.js build output, dynamic import analysis), one investigating the client-side filter/sort hypothesis (Playwright trace, longtask measurements). Have them message each other to disprove competing theories. The theory that survives adversarial scrutiny gets recorded in QA_PERFORMANCE.md as the root cause."

This is the exact pattern the Claude Code docs call out for debugging: parallel investigation with debate. The theory that survives is much more likely to be the actual cause than what a sequential investigator would land on.

What stays in subagent mode

• qa-environment — one-shot setup, no debate needed
• qa-flow-tester — per-feature parallelism; flow-testers for different features have nothing to say to each other
• qa-consistency-checker — single-perspective cross-module audit; teammates would just duplicate work
• qa-performance-prober — runs the systematic measurement pass; the competing-hypotheses team only spawns on specific findings, not as a replacement