What problem does it solve? Agents often report a form submit, write, navigation, fix, or review verdict as successful based only on the action's own return value or a stale pre-action snapshot, producing verification hallucinations. This Skill enforces a claim → fresh ground-truth → confirm loop so every success report is backed by a post-action read from the source of truth. ## Core Features & Use Cases - Claim-first verification loop: State a checkable observable before acting, re-derive it from a fresh snapshot (a11y tree, DB query, file re-read, test re-run, API GET), then confirm or recant with the fresh read as citation. - Stale-ref validation: Detects references (DOM refs, row ids, line numbers) captured before a state change and forces re-resolution by stable identity such as role + name or primary key. - red-browser integration: Uses red-browser snapshot for token-cheap a11y-tree snapshots with numbered refs, monotonic snapshotIds, and console/network capture as browser ground truth. - Use Case: Inside /verify, after clicking a submit button, take a fresh red-browser snapshot and confirm the success banner node actually appeared before reporting the form submission worked. ## Quick Start Use the ground-truth skill to verify my last UI edit actually changed the target element by taking a fresh red-browser snapshot before reporting it done.
