Tuning a Live Big Screen with One Prompt Hinges on Whether the Switcher Console Lets an Agent Take Over

On-site big screens, Line Out, HDMI 2, RD sheets, asset stretching, and cropping—put these words into one Prompt, and it sounds like a director's checklist. The tweet author Robin used Kimi K3.1 Pro to adjust on-site screen equipment; after sending one Prompt, Robin went to eat breakfast, and by the time Robin came back, it was basically all done.

The original Prompt was very specific:

I've connected to the on-site equipment. Please refer to PROJECT.md for debugging.

Use Line Out for audio output; the on-site big screen is HDMI 2.

Please adjust the asset order and mute strategy according to the RD sheet, and decide on the asset stretching/cropping plan.

The guest coser assets will be provided later; for now, fill the main view.

The information density here is not low. The device connection status is there, the project documentation is there, the audio and video output interfaces are specified, and the asset order, mute strategy, and stretch/crop plan all have a basis. The Agent doesn't need to guess what the venue looks like; it only needs to execute according to the project file and the RD sheet.

A switcher console is the device or software that controls on-site screens, audio, and asset playback. The advantage of building your own switcher console is that the interfaces, states, and operating logic are all in your hands. Whether devices can connect, how assets are arranged, which channel is muted, and whether the image is stretched or cropped can all be shaped into something an Agent can read and call. The phrase "Agent-friendly" in the tweet refers to exactly this.

Many demos get stuck not because the model can't write code, but because the on-site environment doesn't let people operate it. Device states can't be read, interfaces have no documentation, operations have no boundaries, project documentation is missing, and the Agent can only stand outside the door. In Robin's tweet, PROJECT.md and the RD sheet play the roles of manual and operation checklist, Line Out and HDMI 2 pin down the output paths, and the asset strategy has a table to rely on. The grunt work has been cleaned up by engineering, and only then does the model get a chance to take on the job.

Going to eat breakfast after one Prompt sounds a bit performative. The prerequisite for that stylish turn away is that someone has cleaned up the interfaces, documentation, and operational boundaries in advance. In the comments, someone replied 'Really nice,' and someone else gave it an outright like. For people building Agents, what's more worth looking at is that switcher console: it doesn't treat the Agent as an add-on; rather, it leaves an entrance for another kind of operator.