Cover & Title

What does Cover & Title do?

Viewers see both the cover and title together in a list, in under a second. The cover is a small image with a few large words; the title is a single line: the title explains what this is, and the cover provides the visual hook. Both complement each other, not repeat each other. Written for the person publishing a video: attract attention, and every word is delivered by the video.

The easiest way to mess this up: the large words on the cover just repeat the title, so together they're still just one sentence; or the large words are so long they become unreadable when scaled down to a small image; or the cover promises something the video doesn't deliver, making viewers feel misled.

1. Gather Materials

1. Use videos_inspect to view the video: name, aspect ratio (width/height), chapters, and documents. If there is no transcript (document of kind speech), transcribe it first and tell the user the transcription is in progress.
2. Read the transcript: export (kind transcript, format md, timestamps true), use jobs_inspect to wait until completed, then read the files in outputs.
3. The user's current requirements: number of sets, cover word limit, style, language, publishing platform, and any already chosen title—find these in the current conversation. If the language is not specified, use the language of the transcript.
4. Other episodes in a series: if the user wants to differentiate or align with previous episodes' titles and covers, use space_related_videos to see what similar content is called.

2. Generate Title & Cover Copy

1. Start with one sentence that states what this video gives the viewer: what problem it solves, what conclusion it draws, or what they'll see. Both the title and the large words should come from this core idea, not from sentence structure.
2. Provide several candidate sets (5 if the user doesn't specify), each from a different angle: direct statement, question, number, contrast, quote, etc. Each set consists of one title plus one line of cover large words.
3. Cover large words must be short (when not specified by the user, keep it within a few words according to the language's convention; for Chinese/Japanese, roughly 8 characters). They should be readable at a glance when scaled down to a small image, and should not repeat the title's exact wording: the title says "what", the large words give the "hook", or vice versa.
4. Numbers, conclusions, and names in the title and cover must be verifiable in the transcript. Avoid exaggerated claims; do not promise anything the video doesn't have.
5. Include a brief reason for each set, mark one as recommended, and ask the user to pick a set, or request another batch, or more sets based on a given example.

Compare (just for illustration):

Bad: Title "Sharing AI Code Review Practice" + Large Words "AI Code Review Practice" (large words repeat the title, neither says why it's worth watching)

Good: Title "Code written by AI tripled; here's how we changed the review process" + Large Words "Humans only review what AI gets wrong" (title says what it is, large words give the hook about the approach, eight words, and the video actually delivers)

3. Generate the Cover

Only generate after the user picks a set—do not generate all sets beforehand: generating costs time and money, and the unpicked sets would be wasted.

1. Where the image comes from: If a video frame index exists, first use videos_match_frames (videoId, text describing the desired cover image, e.g., "person's face, excited expression" or "product close-up") to get a few candidates, then use videos_frames with range (copy the candidate's framesRange) to view the frames. If there is no frame index (it reports MODEL_NOT_INSTALLED or similar), use videos_frames to grab frames at chapter starts and emotionally or action-significant moments, looking for a frame with a clear subject or person and expressive face or action. The frames are raw footage, without subtitles or overlaid elements.
2. Generate: First check capabilities with skills_read on baocut-catalog-media's "Generate Image" section. There you'll find differences between services, prompt order, and what to do if a model can't generate. Use image with prompt, size matching the video's aspect ratio (e.g., 16:9, 9:16, 1:1; prioritize the ratio the user specified; if the model doesn't support that ratio, use the closest supported one and inform the user; if the model doesn't accept size, omit size and specify landscape or portrait in the prompt). Include video and import the result as a candidate asset for this video (do not place it on the timeline). In the prompt describe the subject, composition, and style clearly (person or subject on one side, leaving clean space for large text on the other side). Describe the chosen frame's content. If the model's referenceImages is not null, you may use the chosen frame (saved as an image asset for the video) as images for the model to follow; models that don't accept input images rely on text alone. If the user wants a real-photo cover, tell them they can directly use the chosen frame and add large text in their platform or editor.
3. Large words: Generated images often have incorrect text. In the prompt, request a clean area for large text. If the user wants text on the image, read the image after generation; if the text is wrong, regenerate or inform the user.
4. If the user doesn't specify a count, generate 2 images. Set count to 2. If the model outputs only one image at a time (maxCount is 1), submit twice.

4. Things to Avoid

• Don't repeat the title in the large words; don't make them longer than what can be read at a glance.
• Don't promise anything the video doesn't deliver. Numbers, conclusions, and names must be in the transcript.
• Don't generate every candidate set before the user picks one.
• Don't ignore incorrect text in generated images.

5. Delivery

• In your reply, provide the chosen title and cover large words, along with the other candidates, for easy reuse later.
• Cover images: for each artifactId returned by image and the imported assets, explain which frame or description was used and why it was chosen.
• Mention any information that was not verified; if a platform was mentioned, remind the user to check against that platform's current cover size and text rules.