Break is a user-invoked agent skill that stress tests a single UI component. It reads the component’s props and data, keeps only the scenarios that can occur in production, and renders the real component once per scenario in a single column on a scratch route, so you can scroll through every case.
The page is the report. Each break is noted under its scenario and named in the vocabulary of the sibling skill that owns the fix, and a run where everything survived is a complete result.
What You Get
- A scratch route in your app that renders the real component once per scenario, each labeled, with breaks noted under their label
- A one-line list of scenarios kept and axes dropped, so a wrong inference is cheap to catch
- A findings table listing the scenario, what was observed and the owning skill, with broken scenarios first
- Scenario sets such as empty, one word, several sentences and one unbreakable string, plus zero, one and ten times the realistic item count
- The page left running until you say you are done, and removal of the route and fixtures on request
Common Use Cases
- Checking a button, badge or card against long, empty and translated labels
- Finding how a list component behaves with zero items, one item and a very large count
- Seeing whether a form field survives a 320px container and a very wide one
- Confirming loading, error and disabled states do not collapse or lose contrast
What Makes It Unique
It observes and issues no verdict, and it isolates the component on purpose to test whether it defends itself against worst-case content. It adds no styles, fonts or simulated themes of its own, because a component observed under those is a different component. Dark mode, zoom and reduced motion are left for you to toggle while viewing.
Under The Hood
The skill is a SKILL.md and a scenario reference, with no scripts. The agent writes new files in your project: a scratch route and fixture data that import your real component. It runs your dev server and opens the page in a browser if one is available, otherwise it hands you the URL to check. Nothing is fetched from the web. The harness never connects to live data or production state, and the files are deleted when you ask.