prove-checkslisted
Install: claude install-skill ValentinNikolaev/agent-skills
# Prove a Check Can Fail
Assess whether a green test, build, CI job, smoke check, or rehearsal supports the claim being made from it. This skill verifies the check, not the entire product. Do not edit production code or external systems to manufacture a result.
## Route precisely
- Use when the user asks whether a passing check is trustworthy, when a check may have skipped its subject, or when a regression test is being offered as proof of a fix.
- Use `code-review` for a diff-wide defect review; it may use one bounded negative control without invoking this full workflow.
- Use `monitor-ci` to follow a running GitHub Actions job to completion. Use this skill afterward only if the user asks whether the green outcome actually proves a behavior.
- Do not use to judge whether a search that returned no matches proves absence, or to write tests from scratch.
## Identify the claim and signal
1. State the precise defect or condition the check claims to detect and the observable that would turn it red. If no such condition can be stated, classify the check as unproven.
2. Record the command, revision, configuration, test selection, input fixture, and environment actually used. Read back setup mutations through a path independent of the write that made them; exit status alone does not prove an edit or override applied.
3. Confirm that the target was exercised. Count named test cases or result events when filters or gates can select nothing. Treat `skipped`, cached, and no-op outco