Agent Skills: Black-Box Test Design (EP + BVA)

Black-Box Test Design (EP + BVA)

Goal

Given an input specification (one or more parameters with their domains/constraints), produce the smallest set of test cases that still covers the behavior — using Equivalence Partitioning (EP) for the interior and 2-value Boundary Value Analysis (BVA) for the edges. Black-box only: design from the spec, never from the implementation.

Core principle (the rule this skill encodes)

• Ordered/contiguous domain (numbers, dates, lengths, counts, money) → EP + BVA.
• Unordered/categorical domain (enums, booleans, code sets, free text) → EP only (no neighbors exist, so BVA does not apply).
• The interior of a valid partition is homogeneous → cover it with one representative.
• The risk lives at the edges. Off-by-one errors shift a boundary by exactly one, so for each boundary test the boundary value + its just-outside neighbor (2-value BVA).
• Do NOT add interior near-boundary values (e.g. for 1–100, skip 2 and 99) — they sit in the valid class and catch nothing the boundary itself doesn't. That is robust 3-value BVA; it has diminishing returns and is not the default.
• Internal thresholds are boundaries too. If behavior changes mid-range (tax brackets, shipping tiers, a discount at qty ≥ 10), each threshold gets its own BVA treatment — it is not "interior."

Procedure

1. List inputs. For each parameter, state its domain and constraints from the spec.
2. Classify each domain as ordered or unordered (see core principle).
3. Partition (EP). For each input, derive:
   • the valid partition(s) — inputs the system should accept;
   • the invalid partition(s) — below range, above range, wrong type/format, empty/null, and any other rejected class. Pick one representative per partition (use a nominal interior value for a valid range).
4. Add boundaries (BVA), ordered partitions only. For each boundary, add the boundary value and the value just across it (2-value). Mark which should be accepted vs rejected. Apply to every internal threshold as well.
5. Combine & dedupe. Test set = EP representatives ∪ BVA edge values.
6. Multiple inputs. Default to one-variable-at-a-time: vary one input across its partitions while holding the others at a valid nominal value. Use a decision table only when conditions genuinely interact to produce different outcomes.
7. (Optional) Error guessing. Add experience-based probes the spec doesn't name: empty string, whitespace-only, zero, negative, leading zeros, very long input, special/Unicode characters, duplicate submission. Clearly label these as supplementary.

Output format

A markdown table, plus a one-line note on what was deliberately not tested (and why).

| # | Input(s) | Covers (partition / boundary) | Valid? | Expected result | |---|----------|-------------------------------|--------|-----------------|

Keep the set minimal. If you drop something a naive reader might expect (e.g. interior near-boundary values), say so explicitly so the omission reads as a decision, not a gap.

Worked example — age accepted when 18–60

• Domain: integer, ordered. Valid partition 18–60; invalid partitions <18 and >60; plus non-integer / empty as invalid.
• EP representatives: 40 (valid), 10 (too low), 70 (too high), "abc" (non-integer).
• BVA (2-value) at each boundary: 17, 18 (lower), 60, 61 (upper).
• Combined set: { "abc", 10, 17, 18, 40, 60, 61, 70 }.

| # | age | Covers | Valid? | Expected | |---|-----|--------|--------|----------| | 1 | 40 | valid interior (EP) | ✅ | accepted | | 2 | 17 | lower boundary − 1 (BVA) | ❌ | rejected | | 3 | 18 | lower boundary (BVA) | ✅ | accepted | | 4 | 60 | upper boundary (BVA) | ✅ | accepted | | 5 | 61 | upper boundary + 1 (BVA) | ❌ | rejected | | 6 | 10 | below-range class (EP) | ❌ | rejected | | 7 | 70 | above-range class (EP) | ❌ | rejected | | 8 | "abc" | wrong-type class (EP) | ❌ | rejected |

Not tested: 19 and 59 (interior near-boundary values) — redundant under EP, covered by #1.

Anti-patterns to avoid

• Enumerating many values inside one partition (no added coverage).
• Adding boundary ± 1 inside the valid class (3-value padding like 2/99 for 1–100).
• Applying BVA to unordered/categorical inputs.
• Designing from the code instead of the spec (that's white-box, not black-box).