SuperAPI AI Directory
Partner中文AI newsFavorites
Skillscover-and-title

cover-and-title

Design & creation

View repository

Updated 10-20 09:10

About this skill

Cover and Title

Viewers see the cover and title together in a list, for less than a second combined. The cover is a small image plus a few large words, and the title is one line of text: the title makes clear what this is, and the cover provides the visual and the hook. The two complement each other and do not copy each other. Write for the person publishing the video: be engaging, and make every line delivered by the video.

The easiest ways to do it badly: the large cover text copies the title again, so the two together are still just one sentence; or the large text is too long and simply cannot be read once shrunk into a small image; or the cover promises something the video does not contain, and viewers click in only to find it is not what they expected.

1. Gather materials

  1. Use videos_inspect to look at the video: name, aspect ratio (width and height), chapters, documents. If there is no transcript (a document with kind speech), transcribe first and tell the user that transcription is in progress.
  2. Read the transcript: export (kind transcript、format md、timestamps true), wait with jobs_inspect until completed, and read the files in outputs.
  3. The user's requirements this time: how many sets, the cover text character limit, style, language, publishing platform, and any already chosen title, all can be found in this conversation. If no language is specified, use the language of the transcript.
  4. Other episodes in the series: when the user wants to differentiate from or stay consistent with the titles and covers of previous works, use space_related_videos to see what the several episodes covering similar content are called.

2. Produce titles and cover copy

  1. First use one sentence to make clear what this video gives viewers: what problem it solves, what conclusion it gives, or what is seen. Both the title and the large text start from this sentence, not from phrasing.
  2. Provide several candidate sets (5 sets if the user did not specify), each with a different angle: direct statement, question, number, contrast, quotation, and so on. Each set is one title plus one line of large cover text.
  3. The large cover text must be short (if the user did not specify, follow the habit of this kind of text and keep it within a few words; for Chinese and Japanese, about 8 characters or fewer), readable at a glance even when shrunk into a small image, and must not repeat the title's exact wording: the title says "what it is," and the large text gives "the hook," or the reverse.
  4. Numbers, conclusions, and names in the title and cover must all be traceable in the transcript; do not use exaggerated assertions, and do not promise anything the video does not provide.
  5. Attach one sentence of reasoning to each set, mark one recommended set, and ask the user to choose one set, or to say they want another batch, or to say which set to use for several more sets.

Compare (just for illustration):

Bad: title "AI Code Review Practice Sharing" + large text "AI Code Review Practice" (the large text copies the title, and neither says what is worth watching)

Good: title "AI-written code tripled, here is how we changed the review process" + large text "People only look at what AI is unsure about" (the title says what it is, the large text gives the hook of the approach, eight characters, and the video really does discuss it)

3. Generate the cover

Do this only after the user has chosen a set; do not generate all the sets first: generation takes time and money, and the sets not chosen are wasted work.

  1. Where the visuals come from: when there is a visual index, first use videos_match_frames (videoId，text write the desired cover visual, for example "person's face front-on, excited expression" or "product close-up") to get several candidates, then use videos_frames' range(copy the candidate's framesRange) Take frames to look at; when there is no frame index (it reports MODEL_NOT_INSTALLED and other reasons), use videos_frames to take a few frames at the chapter start and at places with obvious emotion or action, and find a frame where the person or subject is clear and the expression or action has tension. Frames are source footage images and do not include subtitles or overlaid elements.
  2. Generate: First look at the capabilities, use skills_read to read baocut-catalog-media 's "Generate Image"; the differences between services, the prompt order, and what to do when it cannot draw are all there. Use image to give prompt, and according to the video aspect ratio give size(width and height, for example 16:9、9:16、1:1; the ratio the user names takes priority; when the model does not support this ratio, take the closest one it supports and tell the user; when the model does not accept dimensions, do not give size, and specify horizontal or vertical in the prompt), bring video, and import the result as candidate material for this video, without putting it on the timeline. In the prompt, clearly write the subject, composition, and style (the person or subject on one side, leaving the large text in the clean area on the other side), and describe the image of the selected frame into it; when the model's referenceImages is not null, you can use the selected frame (saved into the video's image material) as images for it to draw from, while models that do not accept input images rely only on text; when the user wants a cover with a real person's photo, tell the user they can directly use that frame, and the user adds the large text in the platform or editor.
  3. Large text: Text in generated images is often written incorrectly. In the prompt, require leaving a clean area for the large text; when the user wants text in the image, after generating, read the image once, and if the text is written incorrectly, redo it or tell the user.
  4. When the user does not specify the number of images, generate 2,count give 2; when the model only outputs one at a time (maxCount is 1), submit in two separate times.

4. What to avoid

  • Do not copy the title for the large text, and do not exceed a length that can be read at a glance.
  • Do not promise things the video does not provide; numbers, conclusions, and names must all be in the script.
  • Do not generate every group before the user has chosen.
  • Do not ignore incorrectly written text in generated images.

5. Delivery

  • In the reply, give the title and cover large text the user has chosen, as well as the other candidates, so they can be changed later.
  • Cover image: each artifactId returned by image and the imported material; explain which frame was used or what description was used, and why it was chosen.
  • Which information has not been verified; for anything about the platform, remind the user to check it against the platform's current cover dimensions and text rules.