A diagnosis is a traceable decision, not a magic label.

01 · Candidate formation

Closed content set

The engine matches names, aliases, common descriptions, behavior signals, negative constraints, and cluster relationships. It never invents a component outside the 20 published entries. Internal scores rank candidates; they are not probabilities and are never shown as confidence percentages.

02 · Questions

Every answer must change the decision

A discriminating question is allowed only when an answer can reorder candidates, exclude a candidate, or alter the implementation assumptions. The flow asks no more than three questions and always offers “Not sure”.

03 · Taxonomy

Not every visible label is the same kind of thing

Dialog and Modal Dialog have a subtype relationship. Drawer and Sheet primarily describe presentation. Autocomplete is usually behavior inside an editable Combobox. Notification and Banner describe channels or containers. The product keeps these familiar entry terms while stating what each name cannot guarantee.

04 · Sources

Claim-level references beat source counts

Accessibility behavior prefers platform documentation and W3C guidance. Component-specific presentation uses official library documentation. A source supports only the claim shown beside it. More links do not automatically make a conclusion stronger.

05 · Privacy

Do not paste secrets

The static release processes the description in your browser. Its event adapter never includes raw query text, source code, clipboard content, email addresses, or query-bearing URLs. A future feedback system would require a separate privacy decision.

06 · Known limits

The tool does not know your whole interface

• It covers web UI only.
• It cannot inspect screenshots or running code.
• It cannot know requirements you did not describe.
• Library names vary across ecosystems.
• A general AI assistant may be faster for well-specified tasks.
• Public availability does not prove market demand or repeat use.