zhangtaoran2010/baocut-subtitles-intel

Narrative Mashup

A narrative mashup is a form, not a topic: a voiceover is running, with real moving visuals underneath (interviews, press conferences, screen recordings, news footage, photos). Let the people involved speak for themselves. Tweets, quotes, dates, and numbers are layered on top of the visuals. A news story, a piece of gossip, the origin of a product, a person's story, a trend recap—all can be done this way; a segment can also be inserted into an explainer video or talking-head cut.

Whether this type of video is engaging depends almost entirely on one thing: whether the viewer's eyes have something to do. A card sitting on screen for ten seconds while the narration plays is a slideshow; the same narration with the person's face, the scene from that day, a tweet popping in and out underneath is a video. This skill provides judgment, techniques, and tools, not a workflow template. The order below follows common logic; adapt any step that doesn't fit the current project.

1. Identify whether this project is of this form and align with the user (Sections 1 & 2).
2. Read the source text; list people, events, and questions (Section 3).
3. Source footage and write the script in parallel, feeding each other (Section 4, details in references/sourcing.md).
4. Visuals: something is always in motion under every beat (Section 5).
5. Script, structure, original audio, and sound (Section 6; optional techniques in references/structure.md).
6. Assemble in BaoCut (Section 7, details in references/assembly.md).
7. Quantify, review, and deliver (Section 8).

1. Is This the Right Form?

• A quick test: write down the three most important lines of narration for this project. Ask what visuals should accompany each line. If the answer is "the person saying that line," "the scene from that day," "that tweet," or "the demo of that product," it's this form. If the answer is "an animated model" or "an image that grows with the narration," it's an explainer video (explainer-video). If the answer is "the user talking to the camera," it's a talking-head cut (talking-head-cut).
• When the user says "make a video about this news / this event / this gossip," "make this article into a video," "how did this come about," "find some footage and stitch it together," "documentary," or "recap," default to this form. If unsure, tell the user what you've chosen and why before showing them the script.
• Using a segment in another video: If an explainer video covers a real press conference, or a talking-head cut needs to insert a clip of the subject speaking, or a short video needs a 3-second hook, this is a segment in this form. For the segment only, follow Sections 4, 5, and 7; there's no need to change the entire video's form.

2. Align with the User First

Ask for what's missing; don't ask for what's already explained:

• What is the source text (link, file, a few sentences of context), the language of the final video, and the orientation (horizontal or vertical).
• Duration: If no time limit is given, let the information volume decide. A news story or a piece of gossip: 2–5 minutes. The background of a product or person: 8–20 minutes. Estimate based on narration word count and measured speaking rate (narration Section 1); don't round to a nice number.
• Where does the footage come from: the user's video library, the web, files the user provides. If all three are available, search all three. Before downloading web videos, get user approval (Section 4).
• Note down the user's named style preferences and taboos (e.g., "tight with suspense," "don't make a slideshow," "no background music") and carry them throughout.

3. Read the Source Text, Make a List

Read the entire source text. Create three tables and write them into the working directory:

• People Table: Every named person, their identity, their role in the key moments of the story, and the original quotes from the source text (keep the original language for easy searching in transcripts).
• Events Table: Chronological order, date of each event, possible visual locations (official channel, press conference, news, screen recording).
• Questions: A single central question the entire video answers, plus smaller questions for each segment. A segment without a question is likely just rephrasing the source material.
• Fact-check every fact, date, and quote against its source. Mark any uncertain items in the script. For generated images or reenactments, clearly indicate they are illustrative and never synthesize a real person's voice or likeness.

4. Source Footage and Write the Script in Parallel

The two tasks feed each other: if you find the subject saying something in their own words, change that beat to use the original audio; if you can't find any footage for an event, rewrite that beat or change the approach. The method is in references/sourcing.md. Key points:

• Three sources in order: the user's video library (search across videos by meaning; don't go online if it's already in the library); the web (BaoCut's material search and download; videos need transcription to cut original audio); files and screenshots provided by the user.
• The license of web videos is often unclear: before downloading, list the videos you intend to use (title, channel, link) for the user to decide. Log the source of every piece of footage used and include the list upon delivery. X (Twitter) is read-only: search, browse, read; do not like, retweet, or post.
• Set the in/out points of original audio using the transcript's timestamps to 0.1 seconds, then visually check that the person is speaking in frame.
• If there's a lot of footage, delegate sub-agents to search for people and events by topic; have them write results to files, not in replies.

5. Visuals: Something Is Always in Motion Under Every Beat

This section is the most critical part of the skill.

• The root cause of a slideshow is one thing only: a card (title, quote, bullet points, comparison table) stops on screen for seconds or tens of seconds with nothing moving in the frame. Adding background motion effects or particles won't save it; only real footage underneath will.
• The bottom layer is always moving: underneath every beat of narration, there should be a real video clip, screen recording, or photo (photos should slowly zoom or pan, 3–6 seconds in one direction). Do not leave a solid-color background. If there are truly a few seconds with no footage, make it intentional negative space (a few lines of "thinking" text on a black screen, a blinking cursor), not a card filling the gap.
• Cards sit on top of footage, they do not replace it: Dates, lower-thirds (name/title), quotes, tweets, numbers, terms, chapter titles—these are all overlays. Darken the underlying footage slightly and keep it playing. One card per piece of information, stay on screen just long enough to read through twice (estimate 5–6 characters per second for Chinese). Cards usually cover 20–30% of the screen at most; if they cover more than half, it's likely a slideshow.
• Draw the card as what it is: a tweet is drawn as a tweet (avatar, name, date, original text and translation), a quote is drawn as a quote (who said it, where), a terminal is drawn as a terminal (blinking cursor, text typing), a chat is drawn as a chat. Don't put them all into the same "title + body" box. See motion-graphics for drawing methods.
• Cut every 3–6 seconds, on the word: the moment the narration mentions a name or event, the visual should be exactly that (align using the narration's timestamped words). If a single shot runs for more than ten seconds, either it is self-narrating (original audio) or you should change the angle or push in.
• Don't repeat footage: the same clip should appear at most twice, and the second time must have a new purpose ("look at it again now that we've explained it"). If viewers recognize the same clip, it makes the video look like the editor ran out of material.
• Real footage tells real stories: Use generated images and animated reenactments only illustratively and clearly label them. Darkening, desaturating, or slowing down footage are normal techniques to convey emotion, but do not change the original meaning. Credit sources in a corner and list them at the end.
• For things that can't be filmed, use one or two code-style visuals (flowcharts, scales, timelines, comparisons of items). These are also transparent overlays; don't fill the entire screen and take over the whole segment.
• Add a lower-third on the first appearance of any person. Add a date card on the first mention of any event. Only one information card on screen at any moment. Do not overlay additional cards during original audio clips.

6. Script, Structure, Original Audio, and Sound

• Write the script one beat per line. Beneath each line, on a new line, write what should be visible (visual intent) and where original audio goes. By the time the narration is recorded, you'll know what appears under every beat. Start visual intent lines with @ and skip them during recording. For speaking pace, word count, voice style, and re-recording, follow narration.
• Let the subject speak for themselves: Do not paraphrase the most critical, emotional lines (a decisive statement, an admission, a prediction). Set it up with narration, let the original audio take over, then continue from what the original audio said—don't repeat its content. Each segment: 5–20 seconds, with bilingual subtitles and a lower-third.
• Choose structure based on topic: A quick news item just needs a direct opening. A story benefits from a cold open, a preview, chapter hooks, "thinking" transitions, and time anchors. See references/structure.md for techniques and script writing—these are optional, not steps.
• Sound: Use a low, sparse music bed that rises and falls with chapters and lowers during original audio. Use light sound effects for transitions and card appearances—one consistent set throughout the video. The voiceover should be calm with a hint of intrigue, and the voice style should not change throughout the video.

7. Assemble in BaoCut

Details in references/assembly.md. Key points:

• Track order (bottom to top): real footage (including photos and screen recordings) on the bottom visual track as the base. Place card code packages as a transparent layer above that. Then a watermark. One audio track for the narration, original audio clips attached to their real footage segments (enable their embedded audio), leave the narration silent during those clips. One track for music.
• Record the narration first, align it to get word timestamps, then cut the visuals to those timestamps. If a segment needs to be re-done, re-record and re-assemble only that segment.
• This can be done without a command line: use only the tools (download, transcription, import, edits_apply, code package import, subtitle layer, export). If a command line is available, do the same with baocut CLI and write your own pipeline. This skill does not prescribe the shape of the pipeline.

8. Quantify, Review, and Deliver

• Quantify the timeline (use skills_read to read production-scripts and its coverage.mjs): what percentage of screen time is real footage, code visuals, and blank space; number of shots and median shot length; the longest segment without real footage; any repeated footage; segments with original audio. For this form, expected numbers are roughly: 80–90% real footage, 20–30% cards, median shot length 3–6 seconds, no segment longer than 20 seconds without real footage. If the numbers are far off, it's likely that cards are acting as the base layer—go back to Section 5 to add visuals. The tool only reports numbers, it doesn't judge right or wrong.
• Watch it on mute: does the story work visually on its own?
• Listen without watching: do the joins between original audio and narration feel smooth? Is the music overpowering the voices?
• Check consistency: every date, number, and quote on cards matches the script. The subtitles for every original audio clip are the exact words spoken.
• Deliver: the final video (or review link), the length and chapter markers, a source list for every original audio clip and footage piece (source, link, publisher, in/out points, credits), what is illustrative, any known issues. If a publishing caption is needed, follow titles-and-description.

Reference Files

File	When to Read
references/sourcing.md	Before sourcing footage, setting original audio in/out points, and logging sources
references/structure.md	Before writing the script; when using techniques like cold opens, previews, chapter hooks
references/assembly.md	Before laying out tracks, placing original audio, making cards, and building subtitles in BaoCut

Scripts

This skill comes with no scripts. If a command line is available and narration is recorded beat by beat, use skills_read to read production-scripts for ready-made scripts for dubbing, alignment, timeline assembly, cutting real footage, and measuring visual coverage; it's a convenience path—if the form doesn't match, build it yourself with BaoCut's tools as outlined in Section 7.