Theo Yu 于学龙

This site began with a handover: I wrote down the projects, my experience, the screenshots, and what must never be published, and gave it to an AI agent. It came back with three interactive prototypes — a ring of work, a zoomable map, a desk you can pick things up from — and I chose the ring: a person at the center, the work orbiting around. Earlier rounds, a dark version and a set of static layouts, I turned down; each time I pointed at the real page, said what was wrong, and it rebuilt.

The division of labor is clear. What can go public, which phone numbers become placeholders, which responsibilities are confirmed — those calls are mine. The agent turns them into something that runs and can be checked: a static site, Chinese fonts subset to the characters on the page, content that reads in full without JavaScript, keyboard and screen-reader support, and a build that checks links, metadata and the security policy every time.

The biggest change agents bring is that one person can finish a whole product: prototype, build and verify, several rounds, quickly. But an agent won’t decide what’s worth making or what must stay unsaid. I’ve written those boundaries down as standing rules — for commits, privacy checks and releases — and they are now the most important part of how I work.

TSX starts from the current selection and opens a floating panel with a shortcut; screen capture handles text that cannot be selected. Both paths answer one question: how to stop carrying text into another app.

There are limits. Whether a selection can be read depends on the source app exposing accessible text, and recognition may need correcting. A complete interaction cares for the happy path and states failure clearly — along with what you can do next.

So the result stays in the panel, where both the original and the translation can be edited. A good translation tool should step aside the moment you’re done, and let you get back to work.

Clip puts every kind of clipboard content into one library, with search, types and favorites as ways back in. Tempo centers on the month view, with reminders and world clocks one level away. Both fit very different information into very little space.

The approach: keep frequent actions steady and visible, and reveal rare details on demand. Categories, whitespace and consistent placement orient people better than more decoration.

When an interface quiets down, people find what they came for faster.

Pennilo’s public notes keep recognized drafts separate from saved records. An image is recognized on device and lands as a draft; the user reviews it before saving. That step keeps “what the system saw” apart from “what I chose to record.”

The boundary should show in every state: unconfirmed content should never look saved, failure should never look like success, and discarding a draft should never touch existing records. Clear feedback turns automation into help rather than a new kind of doubt.

It costs one extra tap, and buys a record you’ve actually checked.