2026-05-03

Window into China started as a curated visual portal: city cams, pandas, street food, travel, transit, and creator videos. The Mandarin-learning scan exposed a cleaner second audience: people who want to hear and practice Chinese through culturally situated video.

The useful product shape is not a generic “learn Chinese” bucket. It is a channel ladder.

31 -> easiest Mandarin window
32 -> beginner daily listening
35 -> lower-intermediate listening
38 -> upper-intermediate conversation
41 -> native-speed context
46 -> Mandarin through food, travel, and city life

That keeps the core Window TV model intact. The product still curates embeddable videos, respects rights posture, and routes sources into channels. The difference is that learner fit becomes first-class metadata.

The first fields are deliberately coarse:

language_profile -> english_heavy, chinese_heavy, mixed, ambient, unknown
speech_profile   -> hosted, conversation, ambient, lesson, music, unknown
learner_level    -> none, beginner, intermediate, advanced, native, mixed, unknown

This is enough to start routing without pretending the system can precisely grade every video. Creator labels such as HSK level are hints, not truth. Actual fit depends on speech speed, subtitle scaffolding, topic complexity, format, and whether the content is mostly Mandarin from the viewer’s seat.

The first pass promoted experimental non-default rows for beginner Mandarin, daily listening, and Mandarin Corner interviews. The next pass expanded into intermediate listening, culture podcasts, daily life listening, conversation, news context, and Mandarin through food.

The guardrail is important: learner rows stay experimental and non-default until language-learning visitors show intent. They should not crowd out the ambient China-window use case.

The lesson is product architecture, not just content curation. A channel number can be a promise. Once the promise is “this is the next harder Mandarin rung,” source metadata, QA, and guide layout all need to support progression rather than just topic matching.