如何根據影片逐字稿生成 YouTube 標題和縮圖構思,又不替創作者添加未曾説過的內容
Generating YouTube titles and thumbnail ideas from a transcript without putting words in the creator's mouth
Cutora 根據影片逐字稿生成 YouTube 標題、描述和縮圖構思,並以規則避免替創作者添加未曾説過的內容。標題和描述只取材自逐字稿,描述必須用創作者的第一人稱;縮圖構思則結合影片畫面時間點與創作者當時説過的短句。資料不足時,編輯器會説明原因並不生成章節;所有內容打包成 ZIP,不會代創作者發布。
After the first cut of a talking-head video, creators still need a title, a description and a thumbnail. Language models are good at writing those. They're also good at writing them in a voice that isn't the creator's. Here's how we approached it in Cutora, and the rules that mattered more than the prompt.
Rule 1: the source is the transcript, nothing else
Titles and descriptions are written from what the person actually said in this video. No web facts, no "industry insights". If it isn't in the transcript, it doesn't go in. That keeps the output checkable: the creator can always point to where a line came from.
The side effect is honest: if the creator said something wrong, the draft repeats it. We say so in the UI ("check it before you publish") rather than pretend the model fact-checked anything.
Rule 2: first person, always
Early drafts described "the speaker". Nobody publishes a description of themselves in the third person. The description now has to be in the creator's own voice ("I"), and that's a check, not a style hint.
Rule 3: a thumbnail idea is a timestamp plus the creator's own words
We don't generate images of people. A thumbnail idea is a moment in the video and a short phrase the creator said around that moment. "Make the picture" grabs that frame and draws the words on it, in the browser, as a PNG. The creator's face, the creator's words.
Rule 4: say when you won't
Chapters need enough material. Under a minute, or without three clear topics, the editor makes none and says why, instead of inventing "Intro / Main point / Outro" for a 28-second clip.
Packaging
Everything lands in one ZIP: a README, the publish text, the thumbnail pictures, and any Shorts that exist. Nothing is posted anywhere on the creator's behalf.
What I'd do differently
Most of the work was in rules 1 and 4: stopping the model from being helpful in ways the creator didn't ask for. The prompt was the easy part.
Disclosure: I build Cutora. More on how the first cut itself is decided: https://usecutora.com/guides/how-the-first-cut-decides-what-to-remove?utm_source=devto&utm_medium=pseo&utm_campaign=publish-kit
來源:DEV Community · dev.to