Evidence-Driven Validation
Use this skill when the user wants more than a quick test and explicitly cares about showing that programmed behavior is correct.
Typical triggers:
- "prove it works"
- "don't just eyeball it"
- "show me why this is correct"
- "I want a real validation plan"
- "this transition / workflow / interaction must be exact"
- "I want measurable correctness, not vibes"
Goal
Turn the programmed system into an observable model and then check:
- what the valid states are
- which transitions are allowed
- which invariants must hold during and after every transition
- which artifacts prove the result
Workflow
- Define the state model
- inputs
- internal state
- outputs
- persisted state
- task-specific runtime state such as queue progress, active step, or selected entity
- Define validation obligations
Write them in an if-then form:
- If preconditions hold,
- and transition is applied,
- then invariant set must still hold,
- and the resulting observable state must match the expected state.
- Add diagnostics
Add the smallest observability layer that makes the obligations checkable.
Expose at least:
- current state values
- transition phase
- counters
- derived outputs
- persisted snapshot checksum when persistence matters
- error count or failure signals
- Add deterministic runs
Use automation to run repeatable cases such as:
- idle stability
- transition sequences
- persistence roundtrip
- boundary conditions
- failure handling
- regression cases for bugs that were already found
- Save artifacts
For every validation case, save:
- JSON state snapshot
- pass/fail verdicts
- log output
- screenshot or trace when visuals matter
- optional video when motion or timing matters
- Keep one manual protocol
If a critical behavior cannot be simulated reliably, write a short manual protocol with exact pass rules and required evidence.
Output Standard
The result is only complete when all of these exist:
- diagnostics mode or equivalent observability hook
- deterministic validation runner
- artifact directory
- readable report
- clear list of currently passing vs failing validation cases
Important Rule
Do not claim success from appearance alone.
If the validation runner says a case fails, report it as failing even if the feature looks mostly right.