🧪 QA Roadmap Manual · Automation · AI — an open curriculum by Arifuzzaman Antor

Stage 16 — AI-Powered QA: Become the 100× Tester (the module)

Goal: turn AI from a novelty into your daily force-multiplier — know exactly what to learn before you implement it, what to automate, how to stack tactics into ~100× throughput, and how to do it safely. Stage 11: AI × QA is the hands-on MCP + Playwright + evals primer; this module is the strategy, the prerequisites, and the adoption playbook. Field notes from running AZAI Labs (build with agents, not headcount).

The 100x QA stack: AI leverage per activity compounding to ~100x throughput

1. What “100×” actually means (honest math)

You will not run one magic tool that’s 100× faster. You’ll apply many 3–10× tactics across your day, and they compound: faster test design × instant test data × agent-driven exploration × generated test code × instant log triage. Ten activities each 2–10× faster is a different job, not a faster version of the old one. AI removes the toil; your judgment becomes the scarce, valuable part.

The mental model:

Old QA = do the work + judge the work. · AI QA = direct the work + judge the work at 10× the volume.

2. 🧠 Before you implement AI — what QA MUST know first

Skipping this is why most AI-QA efforts produce confident garbage. AI amplifies whoever’s driving — a strong tester gets 100×, a weak one 100×’s their mistakes. Master these before you automate anything with AI:

Prerequisite Why it’s non-negotiable Where to learn
Test design & risk thinking You must judge AI output — spot the missing boundary, the weak assertion, the wrong oracle. AI proposes; you decide Stage 1, Stage 4
What “correct” looks like (oracles) AI can’t tell you the right answer for your product — you supply the oracle Stage 0
Basic programming + CLI + Git To run agents, read generated code, and commit it Stage 14
How LLMs work (high level) Tokens, context windows, non-determinism, hallucination, temperature — so you know when to trust output and when not to Stage 11
Prompting fundamentals Role + context + constraints + examples = usable output. Vague prompt = vague result §7 below
The stack: MCP, an AI client, Playwright, an eval tool The actual tools you’ll wire together §5 below
Data & security hygiene Never paste secrets, credentials, customer PII, or proprietary code into a prompt unless the tool is approved for it. Know your company’s AI policy Stage 13
Verification discipline AI drafts, you decide. Every AI artifact is a draft until you’ve verified it everywhere

✅ Readiness check — you’re ready to implement AI when you can:

  • Design a strong test suite from a vague requirement without AI
  • Look at an AI-generated test and say precisely why an assertion is weak
  • Run a script from the terminal and commit it with Git
  • Explain hallucination and non-determinism in one sentence each
  • Write a prompt with role, context, constraints, and an example
  • State what data you must never put into a prompt

3. 🗺️ What AI can automate across the QA lifecycle

QA activity AI leverage Tool Your (human) role
Requirement/AC review Flag ambiguity & gaps before coding any LLM Decide what’s actually a gap
Test-case design Generate EP/BVA/negative/edge cases LLM + your prompt Prune, prioritize, add domain risk
Test data Unicode/RTL/boundary/malformed sets instantly LLM Confirm realism & privacy
Exploratory testing Agent walks the app via MCP Playwright MCP + Claude Code/Cursor Set the charter; judge findings
Test code generation Draft Playwright/API specs from a description or recording codegen + agent Review assertions (the job)
Self-healing automation Auto-fix drifted locators Playwright healer, Testim, Mabl Approve/reject the fix
Bug reports Turn messy notes into clean reports LLM + your template Verify repro & evidence
Log / trace / HAR triage Explain stack traces, find root cause fast LLM Confirm the hypothesis
Regression selection “Given this diff, what’s most at risk?” LLM Final risk call
Visual testing AI diffing that ignores noise Applitools, Percy Approve intended changes
Docs & test summaries Draft summaries, coverage notes LLM Fact-check
Testing AI features Build eval sets, LLM-as-judge, red-team promptfoo, DeepEval Own the quality bar

4. ⚡ The 100× speed playbook (stack these)

Each tactic below, with how to trigger it. Do them daily until they’re reflexes.

  1. Instant test design (×5). Paste the story + your draft cases: “Act as a senior QA. What conditions am I missing — boundaries, states, concurrency, failure modes? Return a table.” Keep what survives your judgment.
  2. Instant test data (×10). “Generate 40 signup records covering locale, length, format, unicode/RTL, and injection edge cases, as CSV.”
  3. Agent-driven exploration (×8). MCP: “Open staging, explore checkout with expired cards and stacked coupons, report anything broken.” Review its findings.
  4. Generated test code (×6). npx playwright codegen, or ask the agent to emit a spec — then rewrite assertions to check the right things.
  5. Log/trace triage (×4). Paste the failure: “Likely root cause? What should I check next?”
  6. One-read bug reports (×3). Notes + template → clean report; you verify repro and attach evidence.
  7. Requirement review (×5). “List every ambiguous or untestable statement in this user story.”
  8. Regression targeting. Feed the changelog/diff: “Which of these 40 cases are most at risk from this change?”
  9. Docs & summaries. Draft the test summary report; you fact-check and sign the recommendation.
  10. Evals for AI features. Stand up a promptfoo set so every prompt/model change is scored automatically.

The compounding rule: don’t chase one giant automation. Win 30 minutes here, an hour there, all day — that’s the 100×.

5. 🧰 The modern AI-QA stack (2026)

Layer Tools
AI client (the cockpit) Claude Code · Cursor · Claude/ChatGPT web
Browser control for agents Playwright MCP + Model Context Protocol
Automation engine Playwright (+ its planner/generator/healer agents)
Self-healing / low-code AI Testim · Mabl · Functionize
Visual AI Applitools · Percy
LLM evals (testing AI) promptfoo · DeepEval · Langfuse · OpenAI Evals
AI security OWASP Top 10 for LLM Apps

Master depth, not breadth: one agentic workflow (MCP + Playwright) + one eval tool (promptfoo) beats ten tools used shallowly.

6. 🪜 The adoption ladder (crawl → fly)

Rung You can… Prove it
Crawl Use AI daily for test design, data, and bug reports — verifying every output A week of AI-assisted artifacts
Walk Run an MCP agent to explore a staging app and summarize findings An agent exploration write-up
Run Agent-generate a test suite, harden the assertions, run it in CI A green AI-assisted pipeline
Fly Build eval pipelines for AI features + design an AI-QA workflow your team adopts A team playbook + eval dashboard

7. 📝 Prompt patterns for QA (copy-paste library)

  • Role + constraints: “You are a senior QA engineer. Given this requirement, list the test conditions I’m missing. Focus on boundaries, states, concurrency, and failure modes. Output a table; no prose.”
  • Few-shot (teach your style): paste 2 of your best bug reports, then “Write the next one in exactly this format.”
  • Self-critique (catch weak output): “Here are 10 tests you generated. Now act as a skeptical reviewer and flag which assertions are weak or only check ‘it loaded’.”
  • Grounded (reduce hallucination): “Using ONLY the API spec below, list every endpoint’s negative test cases. If the spec doesn’t say, write ‘UNSPECIFIED’.”
  • Explain-then-act: “First explain the likely root cause of this trace, then propose the next 3 checks.”

8. 🛡️ Guardrails & pitfalls (read before you scale)

  • Hallucinated tests / weak assertions — AI loves asserting “page loaded.” Review every assertion; that review is the QA job.
  • Over-trust — never sign off a release on unverified AI output. AI drafts, you decide.
  • Non-determinism — the same prompt gives different output; pin versions, keep humans in the loop, and eval on every change.
  • Security & privacy — no secrets/PII/proprietary code in prompts unless the tool is sanctioned. Follow your company’s AI policy.
  • Agents in production — staging only, supervised, least-privilege. Auth reuse (storage state), not re-login loops.
  • Flaky agent runs — treat like a new hire’s first week: watched, then trusted.
  • Governance — log what AI generated, who reviewed it, and keep the audit trail.

9. 📊 Prove the speed-up (metrics)

If you can’t measure it, you can’t defend it (or get promoted for it):

Metric What it shows
Cycle time (idea → tested) The headline speed-up
Time-to-first-test for a new feature AI’s biggest early win
Coverage (cases/paths per sprint) More surface, same headcount
Defect escape rate Speed didn’t cost quality (the honest one)
Automation authored per week Agent leverage, real
Flake rate Guardrail on agent-generated tests

10. 🗓️ 30 / 60 / 90-day adoption plan

  • Days 1–30 (Crawl): AI on every test-design, data, and bug-report task. Build the prompt habit + your prompt library.
  • Days 31–60 (Walk→Run): set up Playwright MCP; run agent explorations; agent-generate a small suite, harden it, put it in CI.
  • Days 61–90 (Fly): build one promptfoo eval for an AI feature; write your team’s AI-QA playbook; start measuring cycle time & escape rate.

11. Where this takes your career

AI-fluent QA is the scarcest, best-paid corner of the field — companies find evaluators harder to hire than builders, and QA people are natural evaluators. Roles: AI QA Engineer · AI Evals Engineer · LLM Auditor · Agentic Automation Engineer (see Stage 11 §6 and the Capstone).

✅ Exercise

  1. Run the readiness check in §2 — honestly. Fix any gap before going further.
  2. Pick three tactics from §4 and use them on real work today; note the time saved.
  3. Stand up Playwright MCP and do one agent exploration; review its assertions and keep only the good ones.
  4. Build a 5-case promptfoo eval for any public chatbot feature.

🎓 Want this as a guided, hands-on course with feedback? It’s a track on AZADEMY.

Next → the Capstone: Get Hired as a Sr. QA · or the hands-on Stage 11: AI × QA