ai-test-generation

<objective> LLMs will happily emit fifty plausible-looking tests that assert nothing, target endpoints that do not exist, and duplicate each other. This skill is a staged pipeline that forces structured intermediates — assumptions, coverage matrix, oracle definitions — out of the model BEFORE any test code, so what you get is traceable, reviewable, and grounded in the real codebase instead of ad-hoc generated noise.

Before starting: Check for .agents/qa-project-context.md in the project root. It carries tech stack, test frameworks, naming conventions, selector strategy, and known risk areas that dramatically improve generated test quality. </objective>

Quick Route

The pipeline is the same for every input; only the Step 1 extraction emphasis changes. Jump to the matching row, then run Steps 2-7 unchanged.

Input type	Step 1 extracts	Watch for
PRD / feature spec	Entities, business rules, acceptance criteria, NFRs, stated assumptions	Implicit requirements inferred from "seamless"/"fast" language
User story + AC	Each AC → ≥1 happy + ≥1 negative scenario	ACs that hide multiple behaviors in one line
Code diff (git diff main...HEAD)	New/changed code paths, modified conditionals, removed behavior	Regression scope: test the changed paths, not the whole module
Bug report	Repro steps, expected vs actual, environment	Write a test asserting expected — fails now, passes after fix
OpenAPI / GraphQL SDL	Endpoints, schemas, required fields, enums, auth	Validation, auth-failure, and edge cases per endpoint, not just 200s

Playwright projects also pick an agent integration mode — see Discovery Q2.

Discovery Questions

Check .agents/qa-project-context.md first — if it exists, use it and skip anything already answered there. Then clarify:

1. What is the input source? PRD / spec, user story + AC, code diff, bug report, or API schema. Determines Step 1 extraction emphasis (see Quick Route). For an LLM/AI feature spec, stop — generate eval datasets in ai-system-testing, not Playwright specs.

2. What is the target test framework, and (for Playwright) which agent integration mode?

   • E2E: Playwright (preferred), Cypress. Unit: Jest, Vitest, pytest. API: Playwright APIRequestContext, Supertest, requests.
   • Playwright CLI + agents (recommended for Claude Code / Codex / Cursor): npx playwright init-agents --loop=claude scaffolds planner/generator/healer agents into .claude/agents/ as markdown. They are interactive dev tools that produce standard Playwright tests which run unchanged in CI. Token-efficient; runs inside the agent's loop.
   • Playwright MCP (npx @playwright/mcp@latest): higher overhead, right when the agent must drive a live browser interactively over a long session.
   • Neither — hand-write tests using AI as a scratch-pad helper.

3. What project context is available? Existing test patterns, Page Objects / helpers, data factories / fixtures, CI constraints (timeout, parallelism). More context = less cleanup.

4. What is the review workflow? Full pipeline → human review → merge (default); scenarios only → human writes code; or code → human refines iteratively.

5. What domain knowledge is needed? Regulated industry (healthcare, finance) compliance, domain invariants (money never negative, appointments cannot overlap), known risk areas from past incidents.

Core Principles

1. Pipeline before code. Never generate test code before establishing what to test, why, and how to verify it. The seven-step pipeline exists to prevent premature code generation that targets the wrong things.

2. Structured intermediates are the product. The assumptions document, coverage matrix, and oracle definitions are more valuable than the test code itself. They are reviewable, traceable, and reusable.

3. Separate what from how. Scenario generation (what to test) and oracle design (how to verify) are distinct cognitive tasks. Mixing them produces scenarios biased toward what is easy to assert, with assertions tacked on as afterthoughts.

4. AI generates the first draft; a human reviews and refines. Never ship AI-generated tests without human review. The AI accelerates — it does not replace judgment.

5. Context is everything. Feed the LLM your conventions, existing patterns, selector strategy, and data setup. The more context, the less cleanup.

6. Quality over quantity. Each test has a maintenance cost. Focus on critical paths, complex logic, and known risk areas — not test count.

The Pipeline

Mandatory workflow — agents MUST follow this order:

Step 1: Extract   → Requirements, entities, business rules from input
Step 2: Analyze   → Risks, invariants, edge cases, ambiguities
Step 3: Map       → Coverage matrix (requirement → scenario → priority)
Step 4: Generate  → Candidate scenarios (happy + boundary + negative + security + a11y)
Step 5: Design    → Assertions and oracles SEPARATELY from scenarios
Step 6: Code      → Test code (only after all above exist)
Step 7: Review    → Human review with traceability back to source

Full prompt templates for every step (extraction, risk analysis, scenario, oracle, code) live in references/prompt-patterns.md. Below is the shape of each step's output.

Step 1: Extract Requirements and Entities

Parse the input into structured elements: Entities (with roles/states/attributes), Business Rules (numbered), Explicit Requirements ([REQ-N], stated in source), and Implicit Requirements ([IMP-N], inferred — flag every one for human confirmation). Separating explicit from inferred is the rule that prevents testing assumptions as if they were specifications.

Step 2: Risk Analysis and Invariants

Derive what can go wrong, what must always be true, and where the source is silent.

• Risks — table of Risk | Likelihood | Impact | Source Requirement (e.g. race condition on stock decrement, email delay > 30s).
• Invariants (must ALWAYS hold) — stock >= 0, order total = sum(items) + tax + shipping, user sees only their own orders.
• Ambiguities (need human answers) — "Does free shipping apply before or after discount codes?" Capture these explicitly; do not silently pick one.
• Edge cases derived from risks — two users buy the last item, payment succeeds but email service is down.

Step 3: Coverage Matrix

The single most important artifact — it prevents both gaps and duplicates. Map every requirement to scenarios with category, priority, and oracle type:

Requirement	Scenario	Category	Priority	Oracle Type
REQ-1	Add single item to empty cart	Happy path	P0	State: cart count = 1
REQ-1	Add out-of-stock item	Negative	P0	UI: error message, cart unchanged
REQ-2	Complete checkout with valid card	Happy path	P0	State: order created, stock decremented
REQ-2	Two users checkout last item	Race condition	P1	One succeeds, one gets stock error
INV-1	Stock never goes negative	Invariant	P0	Data: stock >= 0 after any operation

After building it, verify: every requirement has ≥1 happy and ≥1 negative scenario; every invariant has a direct test; every Step-2 risk has a scenario; no two rows test the same thing.

Step 4: Generate Candidate Scenarios

For each matrix row, write the full scenario in Given/When/Then with explicit test-data requirements (Given: user with 99 items in cart (max 100); When: adds one more; Then: count = 100). Cover these categories systematically:

Category	Description
Happy path	The user does exactly what the feature is designed for
Boundary	Edge of valid input ranges — use the BOUNDARIES framework (references/prompt-patterns.md)
Negative	Invalid inputs, unauthorized actions
Security	Auth bypass, injection, privilege escalation
Accessibility	Screen reader, keyboard-only, contrast
State transition	Valid and invalid moves between states
Concurrency	Two users acting simultaneously

Step 5: Design Assertions and Oracles

Deliberately separate from Step 4. Scenarios describe behavior; oracles describe how to verify it. For each scenario, define oracles across categories — a single assertion is rarely enough to prove a behavior:

Oracle category	Asserts	Example
UI state	Visible text / element state	cart badge toHaveText('1')
Data	Persisted state via API/DB	GET /api/cart returns 1 item, correct total
Negative	What should NOT happen	no error toast; no navigation away
Side effect	Async/external outcomes	analytics add_to_cart fired; email in inbox < 30s

Oracle quality rules: assert business outcomes not implementation details; use the most specific assertion available (toHaveText('$29.99'), not toBeTruthy()); include negative assertions; verify data integrity, not just UI; assert accessibility (focus management, live-region announcements).

Step 6: Generate Test Code

Only after Steps 1-5 produce reviewed artifacts. Code is a mechanical translation of scenarios + oracles into framework syntax, with traceability comments linking back to the requirement and scenario:

/**
 * Scenario: SC-001 — Add single item to empty cart
 * Requirement: REQ-1 (User can add items to cart)
 * Priority: P0
 */
test('add single item to empty cart', async ({ page, testProduct }) => {
  await page.goto(`/products/${testProduct.id}`);                       // Given
  await page.getByRole('button', { name: 'Add to cart' }).click();      // When
  await expect(page.getByTestId('cart-badge')).toHaveText('1');         // Then
  await expect(page.getByTestId('error-toast')).not.toBeVisible();      // Negative oracle
});

Code generation rules: match project conventions (from qa-project-context.md); reuse existing Page Objects, fixtures, and data factories; include traceability comments (Scenario: SC-XXX, Requirement: REQ-XX); follow the project's selector strategy; put setup/teardown in fixtures, not inline.

Step 7: Human Review

Not optional — a mandatory pipeline step. This reviews the tests this pipeline just generated, before they merge. (To audit a pre-existing suite you did not just generate, use ai-qa-review instead.) Run every generated test against this checklist:

• Traces to requirement: test → scenario → coverage row → requirement is followable.
• Tests behavior, not implementation: survives a harmless refactor.
• Correct abstraction level: right test type (unit vs integration vs E2E).
• Test naming and readability: the test name states the behavior; a reader sees intent without decoding the body.
• Test isolation / no shared state: the test creates and cleans up its own data, holds no order dependency on sibling tests, and passes when run alone or in any order.
• Realistic test data: plausible, diverse, using example.com.
• Meaningful assertions: matches the oracle definition; specific, not toBeTruthy().
• Matches project conventions: naming, structure, selector strategy.
• No flakiness risks: no hardcoded timeouts, race conditions, or order dependence.
• Edge cases included: goes beyond the happy path.
• Assumptions validated: Step-2 ambiguities were resolved before coding.

Review outcome per test: KEEP (merge as-is) · MODIFY (fix listed issues, then merge) · REJECT (wrong requirement, wrong abstraction, hallucinated API) · DEFER (blocked on ambiguity).

Guardrails

Hard rules. Agents MUST follow them.

• Code before coverage is forbidden. Never emit test code before Steps 1-3 (requirements, risk analysis with documented assumptions, coverage matrix) exist. If an agent skips to code: STOP, go back.
• Assert outcomes, not implementation. expect(screen.getByRole('progressbar')).toBeVisible(), not expect(component.state.isLoading).toBe(true). expect(page.getByTestId('cart-badge')).toHaveText('1'), not expect(store.dispatch).toHaveBeenCalledWith(...).
• Scenarios (Step 4) before oracles (Step 5), always. Scenario = WHAT happens; oracle = HOW to verify. Mixing them biases scenarios toward easy assertions.
• Always produce the intermediates — assumptions document, uncovered ambiguities, oracle candidates, and the traceability chain — even in abbreviated form.

Flag these when detected:

• Hallucinated APIs — endpoints, selectors, methods, or imports that do not exist in the codebase. Verify mechanically (see Verification) before human review.
• Duplicate scenarios — same behavior, trivially different data. Consolidate or parametrize.
• Low-value assertions — expect(response).toBeTruthy(), expect(page).toHaveURL(/.*/).
• Missing negative cases — if every scenario is a happy path, the coverage matrix is incomplete.
• Unrealistic test data — test@test.com, John Doe, password123. Use diverse, plausible data on example.com.

Model selection per step

Route by difficulty, not habit. Use a cheap model for mechanical extraction (Step 1) and the coverage-matrix bookkeeping (Step 3) — Haiku 4.5 or Sonnet 4.6 are plenty. Escalate to Opus 4.8 for oracle design (Step 5) and hallucination-sensitive code generation (Step 6), where a wrong inference is expensive; reach for Fable 5 only on genuinely hard reasoning (subtle invariants, regulated-domain logic). Running the strongest model on every step is wasteful; running the cheapest on Step 6 produces fabricated APIs.

Verification

Convert the "hallucinated APIs" warning into a mechanical gate. After Step 6, before human review:

1. Resolve imports / types. TypeScript: npx tsc --noEmit — fabricated imports and wrong signatures fail here. Python: python -m pyflakes <files> or ruff check.
2. Grep generated selectors/endpoints against the codebase. Confirm every getByTestId('...') id and every API path the test calls actually exists in source:

   grep -roE "getByTestId\('([^']+)'\)" generated/ | sed -E "s/.*'([^']+)'.*/\1/" | sort -u \
     | while read id; do grep -rq "$id" src/ || echo "MISSING testid: $id"; done
   

3. Run the suite once. Tests that reference nonexistent routes/selectors fail fast; quarantine those before review rather than reviewing dead code.

Any MISSING line or tsc error is a hallucination to fix before a human spends review time.

Anti-Patterns

1. Skipping to code. The most common failure: an agent gets a PRD and immediately writes tests. Without the coverage matrix it misses scenarios and duplicates others. The pipeline exists to prevent this.
2. Asserting implementation instead of behavior. expect(component.state.isLoading).toBe(true) breaks on any refactor. Assert expect(screen.getByRole('progressbar')).toBeVisible() — what the user observes.
3. Mixing scenarios and assertions. Writing "test this thing and check this value" as one step. Separate what to test from how to verify it.
4. No project context in the prompt. Without conventions and existing patterns, you get generic tests. qa-project-context.md exists for exactly this.
5. Over-generating. AI will write 50 tests for a simple function. Each carries maintenance cost. Use the coverage matrix to bound generation to meaningful scenarios.
6. Copy-paste without understanding. If you cannot explain what a generated test does and why, do not merge it. Tests you do not understand become tests you cannot debug.
7. Shipping without review. Step 7 is not optional. AI tests routinely contain hallucinated APIs, wrong selectors, incorrect business logic, and flakiness only human review catches.
8. Ignoring the feedback loop. When AI tests catch real bugs, note the prompt patterns that worked; when they false-positive, note what went wrong. Build a project-specific library of what works.

Done When

• All seven artifacts exist: requirements document, risk & invariants, coverage matrix, scenario set, oracle definitions, test code, and review notes with a KEEP/MODIFY/REJECT/DEFER decision per test.
• The coverage matrix was produced and reviewed before any test code file was written.
• Verification passed: tsc --noEmit (or language equivalent) exits 0 and the selector/endpoint grep reports zero MISSING lines.
• Each generated test has a recorded human review decision; no test is marked KEEP without one.
• The suite's CI job exits 0 (green).
• Reproducibility metadata recorded: the exact model ID (e.g. claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5-20251001), input source hash, and the version of any skill / CLI / MCP server invoked.

Related Skills

• qa-project-context — Set up the context file that makes AI test generation dramatically better. Configure this first.
• playwright-automation — Deep Playwright patterns (POM, fixtures, CI), plus the Test Agents (init-agents --loop=claude, scaffolded into .claude/agents/) and @playwright/mcp modes chosen in Discovery Q2. Generated tests live inside this framework.
• unit-testing — Jest, Vitest, pytest patterns for unit-level generated tests.
• api-testing — Endpoint test patterns for tests generated from OpenAPI specs.
• test-strategy — Decide what to test and at which level before generating.
• test-reliability — Make generated tests reliable: flake classification, healing, video receipts.
• ai-system-testing — When the input is an LLM feature spec, generate eval datasets here (Promptfoo, DeepEval, Ragas, Braintrust) instead of Playwright specs.
• ai-qa-review — Audit a pre-existing suite you did not just generate (test smells, testability). Step 7 here only reviews this pipeline's own output.
• ai-bug-triage — When generated tests find bugs, classify and report them through the triage pipeline.

Reference Files (in references/)

• prompt-patterns.md — Full prompt library aligned to the seven steps: extraction, risk analysis, scenario generation, oracle design, and code generation prompts, plus the BOUNDARIES edge-case framework used in Step 4.