Skip to content

description: Full end-to-end AI test generation loop — 10-step workflow, plan review checklist, code review checklist, and reviewer assignment guide. tldr: The full loop is 10 steps: state intent, gather inputs, invoke Planner, human reviews plan, commit plan, invoke Generator, human reviews code, run tests, commit on pass (or debug/Healer on fail), then CI runs with Healer on locator drift only. Three separate reviewers for plan, generated code, and every Healer patch — same-person review of all three defeats every gate.

End-to-End Workflow

When to Use

Use this when running the full loop for a new feature from intent to CI.

Pattern: 10-step workflow

1. State intent          → "Test the contact form per JIRA-1234"
2. Gather inputs         → user story + code paths + style reference
3. Invoke Planner        → produces specs/contact-form.md
4. Review plan (human)   → PM/dev/QA reads; approves/edits
5. Commit plan           → git add specs/contact-form.md
6. Invoke Generator      → produces tests/contact-form.spec.ts
7. Review code (human)   → developer reads; tweaks if needed
8. Run tests             → npx playwright test contact-form
9. If pass → commit code → git add tests/contact-form.spec.ts
   If fail → debug:
   - Plan wrong? Edit plan, regenerate code
   - Site bug?   Open ticket, mark test fixme
   - Locator drift? Invoke Healer
10. CI runs forever      → Healer on locator drift only

Pattern: review checklist for plans

  • [ ] Title is a single imperative scenario
  • [ ] Preconditions describe state only, not procedure
  • [ ] Steps are 3–10 numbered behavioral actions
  • [ ] Each expected result is observable
  • [ ] Negative checks are present and specific
  • [ ] Out-of-scope is explicit
  • [ ] No selectors in plan
  • [ ] No hallucinated fields (cross-check against live site or code)
  • [ ] No invented acceptance criteria — every bullet traces to intent
  • [ ] Clarifications block resolved before approval

Pattern: review checklist for generated code

  • [ ] Each test() maps to one scenario in the plan
  • [ ] Each test.step() maps to one numbered step
  • [ ] Each expect() maps to one bullet under Expected/Negative
  • [ ] Selectors use getByRole / getByLabel / getByTestId — not class chains
  • [ ] No page.waitForTimeout(...) — auto-wait or waitForResponse
  • [ ] Helpers and fixtures used where appropriate (not inlined)
  • [ ] // spec: and // seed: comments trace back to source artifacts

Decision: who reviews what

Artifact Reviewer
Plan PM / designer / QA / dev — anyone who understands the intent
Generated code Developer — checks technical correctness
Healer patch Developer — checks locator stability

Common Mistakes

  • Wrong: Same person writes the plan, generates the code, and approves both → Right: no real review gate
  • Wrong: Plan committed without running the Generator → Right: plan reviewed but not validated to be code-generable
  • Wrong: Tests committed without re-running → Right: Generator output not verified to actually pass

See Also