Pairwise Test Design
identify interactions that need at least pairwise coverage after factors, values, and constraints are explicit. Produce PWT-## findings. This Skill organizes traceable test-design candidates only; it does not execute tests or turn a design inventory into coverage, pass, or release evidence.
When to Use
- When you need Pairwise Test Design candidates from test factors, values for each factor, combination constraints, platform/role dimensions, risk evidence, and existing combinations.
- When you need selection rationale, applicability constraints, evidence gaps, and the smallest validation action.
- When inputs are incomplete but a bounded first pass can preserve blocked or unassessed boundaries.
Do not use it to execute tests, invent rules, replace a complete strategy, or accept risk for a Human.
Output Format Options
- Use Markdown by default; use tables, JSON, or CSV only when explicitly requested or required by the delivery format.
- Separate static analysis, unexecuted work, evidence states, and Human decisions; keep items unassessed, blocked, or NOT_RUN when runtime evidence is absent.
How to Use
- Read prompts/pairwise-testing.md and provide the objective, scope, material, environment, and evidence.
- Complete known, missing, conflicting, stale, out_of_scope, and assumptions before findings.
- Record PWT-## with factors, value pair, validity constraint, interaction risk, coverage rationale, source evidence, priority, and validation method, source, evidence state, impact, owner, close condition, and validation.
- Preserve conflicts, unknown constraints, and open questions.
Core Constraints
- do not describe pairwise coverage as all-combination coverage, ignore exclusions, or invent factors or values from experience.
- File presence, names, design declarations, and Eval configuration are not runtime evidence.
- Mark unknowns unassessed, blocked, or pending clarification instead of filling them with convention.
- Do not edit requirements, code, test assets, or target systems.
Pre-delivery Check
- Recorded the six-part input audit.
- Every PWT-## has source, minimum evidence, impact/priority, owner role, close condition, and validation.
- Facts, inferences, recommendations, unexecuted work, and Human decisions remain separate.
- Findings are not full cases, execution results, coverage proof, or release claims.
Reference Files
- Read evals/eval.yaml and matching cases for regression; configuration does not prove project results.
- Use evals/trigger-prompts.csv and evals/local-rules.json for trigger checks; missing skill.selection evidence is BLOCKED.
Common Pitfalls
- Do not turn a method name, file presence, or candidate count into test execution, coverage, pass, or release evidence when scope or evidence is incomplete.
- Do not fill in missing rules, thresholds, data, environments, or results from convention; preserve unassessed, blocked, and pending items.
- Do not expand this specialist design or review into a complete strategy, full test cases, runtime execution, or a release decision.
Best Practices
- Complete the six-part input audit before selecting the smallest traceable and verifiable finding scope.
- Keep the source, evidence state, impact/priority, owner role, close condition, validation method, and residual risk for every finding.
- Write validation suggestions as next actions; do not upgrade package structure, candidate counts, or local Eval configuration into real quality conclusions.