The Edge Case Test Generator: Turn a Feature Description Into a Complete Test Matrix Before You Write Code
Why this prompt matters
Engineering teams that start coding directly from a short ticket typically discover 30-40% of their real edge cases only after a customer hits one in production. On a feature that touches money and inventory, each edge case that ships unhandled turns into a support escalation, an engineering fire drill, a reconciliation headache for finance, and in the worst cases — like the double-refund or stale-pricing scenarios above — direct, hard-to-reverse financial loss. Finding these gaps on paper during planning costs minutes; finding them in production after a customer or auditor does costs hours of firefighting plus the trust damage of a visibly broken checkout flow.
What we use it for
A backend engineer is about to implement a new bulk order cancellation endpoint for an e-commerce API — one function call that touches payment refunds, inventory restocking, and customer notifications at once. The ticket describes the happy path in two sentences. Before writing a single line of code, the engineer runs this prompt against the feature description to surface every edge case, race condition, and integration failure the implementation needs to handle.
Prompt
Role: Act as a senior software engineer and QA architect with deep experience in test design and failure-mode analysis for production systems. Context: I am about to implement the following feature or function, and I want a complete picture of what could go wrong before I write any code. Feature description or function signature: [PASTE FEATURE DESCRIPTION OR FUNCTION SIGNATURE HERE] Language/framework: [E.G. "PYTHON WITH FASTAPI", "TYPESCRIPT WITH NODE.JS"] Edge cases I've already planned for: [LIST THEM, OR WRITE "NONE YET"] Domain constraints or sensitivities: [E.G. "TOUCHES PAYMENT PROCESSING", "MUST BE IDEMPOTENT", "RUNS ACROSS MULTIPLE WORKERS", OR "NONE"] Task: Produce a complete test matrix for this feature, organized into these six categories: 1. Happy path - the 1-3 cases representing normal, expected usage 2. Boundary conditions - empty inputs, maximum sizes, zero/negative values, type edge cases 3. Invalid/malformed input - what a malicious or confused caller might send 4. State and concurrency edge cases - what happens if this runs twice, runs concurrently, or runs when upstream state changes mid-execution 5. Integration failure points - what happens when each external dependency (database, API, queue, cache) times out, errors, or returns a malformed response 6. Silent failure risks - cases where the code might succeed technically but produce the wrong result without throwing an error For each test case, give: a one-line description, the specific input/state that triggers it, and the expected correct behavior. Constraints: Do not give generic testing advice - every case must be specific to the actual feature described above. Flag the 3-5 cases you consider highest risk if left untested, with a one-sentence reason why. If the feature touches money, user data, or irreversible actions, call that out explicitly and treat those paths with extra scrutiny. Output Format: Present the test matrix as a numbered list within each of the six categories above. After the full matrix, add a short "Top 5 Risks If Skipped" section ranking the highest-priority gaps.
Result
Feature: Bulk order cancellation endpoint (cancels up to 50 orders in one call, triggers payment refunds, restocks inventory, sends customer notifications).
1. Happy Path
- Cancel 3 valid, unshipped orders belonging to the requesting user → all 3 refunded, restocked, and notified successfully.
- Cancel a single order via the bulk endpoint with a 1-item array → behaves identically to the single-cancel endpoint.
2. Boundary Conditions
- Empty order ID array submitted → should return a 400 with a clear message, not a silent no-op.
- Exactly 50 order IDs (the stated max) → should succeed; 51 should be rejected with a specific "limit exceeded" error, not a generic 500.
- Duplicate order ID appears twice in the same request → should be deduplicated and refunded once, not twice.
3. Invalid/Malformed Input
- Order ID belonging to a different user → must be rejected per-item with a 403 in the response, not fail the entire batch or silently skip it without telling the caller.
- Order ID that doesn't exist → per-item 404, batch should still process the valid remaining items.
- Order already shipped or already cancelled → reject with a specific status explaining why, don't attempt a double-refund.
4. State and Concurrency
- Same cancellation request submitted twice in rapid succession (double-click/retry) → must not issue two refunds for the same order (idempotency key required).
- Order gets shipped by the warehouse system in the few hundred milliseconds between validation and the cancellation transaction committing → must detect the state change and reject that specific item rather than cancelling a now-shipped order.
5. Integration Failure Points
- Payment refund API times out after inventory has already been restocked → must not leave the order in a state where inventory was returned but the customer wasn't refunded; needs compensating transaction or saga pattern.
- Notification service is down → cancellation and refund must still complete; notification failure should queue for retry, not roll back the whole operation.
- Inventory service returns a malformed response (missing SKU field) → must fail that item explicitly rather than assuming restock succeeded.
6. Silent Failure Risks
- Refund amount calculated using stale cached pricing data instead of the price actually paid → technically "succeeds" but refunds the wrong amount.
- Partial batch success returns HTTP 200 without clearly itemizing which orders failed → caller assumes everything succeeded when some didn't.
Top 5 Risks If Skipped
- Double refund on retry - direct financial loss, highest priority, very likely to occur in production given client-side retry logic.
- Refund/restock partial failure (no compensation) - creates permanently inconsistent state between payment and inventory systems.
- Race condition with warehouse shipping - customer receives a cancelled-but-shipped order, becomes a support and trust issue.
- Wrong refund amount from stale pricing - silent, hard to detect, directly costs money either to the business or the customer.
- Ambiguous partial-success response - causes downstream systems or the frontend to misreport order status to the customer.
Most engineers don't skip edge cases because they're careless. They skip them because a feature ticket describes the happy path in two sentences, and the brain fills in the rest by imagining the feature working correctly. This prompt exists to force a different starting point: instead of asking "what should this do," it asks "what are the specific ways this breaks," and it structures that question so the answer is actually usable as a test plan rather than a vague list of worries.
Why six categories instead of one open-ended list
A plain "list edge cases for this function" prompt tends to produce a handful of obvious null-check scenarios and stop there, because nothing forces the model to look at the harder failure modes. Splitting the request into happy path, boundary conditions, invalid input, state/concurrency, integration failure, and silent failure risks works because each category targets a different root cause of production bugs. Boundary conditions catch off-by-one and type errors. Concurrency catches the bugs that only show up under real traffic, not in a single manual test. Integration failure points catch the assumption that a database or payment API will always respond cleanly. Silent failure risks — cases where the code runs without erroring but produces a wrong result — are the hardest category to think of unprompted, and often the most expensive to discover later, which is exactly why the prompt calls them out as their own category rather than letting them get lost inside "other."
Why the prompt demands specificity, not generic advice
Generic testing checklists ("test null inputs," "test empty strings") are available in any QA textbook and add little value. The constraint that "every case must be specific to the actual feature described" is what turns this from a testing-theory lecture into an actual plan for the function in front of you — it forces the model to reason about what inventory restocking, refund issuance, and notification dispatch specifically do to each other when they happen out of order or partially fail, rather than listing abstract categories without content.
Why the output ends with a ranked risk list, not just a flat matrix
A 20-30 item test matrix is useful for a QA pass, but not every item carries equal consequence, and a busy engineer facing a deadline needs to know which 3-5 cases are non-negotiable even if the rest get deferred to a follow-up ticket. Ending with a "Top 5 Risks If Skipped" section, each with a one-sentence reason, turns an exhaustive list into a triage tool — the difference between a document nobody reads before shipping and one that actually changes what gets built first.
Who should use this, and when
This prompt earns its keep most on features with real consequences for getting an edge case wrong: anything touching payments, inventory, authentication, or data that can't be easily corrected after the fact. Run it at the point you'd normally start writing code, feed it the real ticket or function signature, and treat the output as a checklist to work through — or to hand directly to whoever is writing the tests.