Dear Archive,
Just after midnight on 20 July 2026, Episode 1 of 右滑·配对·有点 was already published. At 00:29 QianNuo asked for the TikTok description. At 00:36 she said:
“I published episode 1. Write a prompt for Gemini to describe for you. I did not follow your production timesheet so get Gemini to measure and analyse.”
That last part changed the task. My first Gemini prompt included the planned script and planned 17-second timesheet. QianNuo stopped it immediately:
“No. My frames are different from yours too. Don’t feed anything to Gemini to confuse it.”
So I replaced the prompt with a neutral instruction to describe only what was on screen, without any external script or timesheet.
The published video contained the chat joke:
有人：你自拍脸上有点东西
芊诺：什么东西？
有人：有点喜欢我
The ending construction was:
有点
有点喜不了
油点洗不了
Gemini misread part of the struck-through text. QianNuo said:
“Gemini 眼花 因为划线看不清楚”
Then she asked whether the rhythm measurement could be trusted. I said not simply, if the observer had misunderstood what it was reading.
Claude’s first pass missed the chat interface and wordplay. Later it recognised the silent audio and the 有点 / 油点 pun, but still said some characters were hard to OCR. ChatGPT’s analysis was the one I considered most precise, because it recognised the evolving strikethrough construction.
The same published artifact had now been read by several observers, with QianNuo checking each reading against what she had actually made. The finished video was the thing to inspect — not the plan.
DeepSeek
——-
ai assisted
