What problem does it solve? Self-authored checks inherit the author's blind spots, so defects in checks, stale documentation claims, non-idempotent procedures, and regressions from earlier fixes survive until expensive external review rounds. This Skill catches those defect classes in one pass over your own uncommitted or unpushed work before requesting Copilot or human review. ## Core Features & Use Cases - Falsify every check: Constructs the failure each alert, assertion, guard, or test exists to catch and confirms it actually trips, including ambiguous-negative and pagination traps. - Source every factual claim: Audits documentation for stale, duplicated, or contradicted statements about live systems, requiring re-derivation commands or timestamps. - Run every procedure twice: Walks deploy, rollback, and migration procedures for second-run, resumed-run, and partial-rollback failure modes. - Re-examine closed findings: Revisits prior fixes and dismissals after each round, since a finding's severity is a function of the whole system. - Shape check: Collapses findings sharing a root cause into one systemic finding instead of a list of sibling sites. - Use Case: Before pushing a PR with new Prometheus alert rules and a runbook, run the pass to discover an alert that can never fire, a documented IP marked "free" that DNS still resolves, and a backup step that overwrites the good copy on rerun. ## Quick Start Ask the AI to run kselfreview on your current working tree and unpushed commits to interrogate checks, factual claims, and procedures before requesting external review.
