What problem does it solve? Design-system rules are written down but still get broken in code, because reading a rule and applying it are different acts and nothing checks the second one. Barb closes that gap by reviewing a screen built on the Recursica design system against the rules the system actually states, reporting every violation with a file and line. ## Core Features & Use Cases - Manifest-driven skill selection: Runs scripts/screen-skill-manifest.mjs to compute which of the 62 Recursica skills apply to a screen based on its actual adapter imports, following local imports and reporting anything it cannot map. - Parallel per-skill checking: Dispatches one checker subagent per applicable skill to walk that skill's pre-flight checklist item by item, since the full corpus does not fit in one context. - Adversarial verification: Dispatches a feisty subagent per finding to argue it is wrong, defaulting to refuted when uncertain, so only confirmed violations reach the report. - Fix re-verification loops: Re-checks fixes against the whole rule rather than the reported instance, looping until two consecutive rounds find nothing new. - Use Case: After building or changing a screen on @recursica/mantine-adapter, point Barb at the route and the knowledge checkout; she returns a list of rule violations with skill, checklist item, file, and line, plus an honest list of what source-only review cannot check. ## Quick Start Ask the agent to review the screen at a given route file against the Recursica skills corpus and report every rule violation with its file and line.
