XB Video Cut v2
This is a speech-led video editing specification for ChatCut. It separates "deciding what to cut" from "what order to place on the timeline": first determine the content goal, then arrange all visual, text, and audio layers according to the final spoken-word timeline.
1. Scope and Basic Constraints
Applicable to existing footage for single-person talking heads, interviews, video podcasts, courses, tutorials, and explainer videos. When the user wants to create a talking-head from scratch without source material, do not use this skill's editing workflow; first state that existing video footage is needed.
Source material may already be a project asset, a local path, or a conversation attachment:
- Material already in the project should be used directly, without re-importing.
- For local readable paths or conversation attachments, register the asset into the project using the current ChatCut Desktop asset registration method, then start reading the transcript.
- If the editor does not have the target project open, ask the user to open it first; do not assume the current project is the target one.
Consistent Chinese product terminology:
A-roll: Voice editingBackground music: Background musicCaptions: SubtitlesB-roll: Always keepB-rollMotion Graphics: MG animationAI Voice Isolation: AI voice isolationTranscript panel: Transcript panel
Do not substitute "motion effects", "subtitle bars", or "dynamic subtitles" for "MG animation", and do not translate "B-roll" into other Chinese terms.
2. First Determine the Goal, Then Choose Processing Items
At each start, confirm these three dimensions:
- Goal: Platform, final video duration, landscape/portrait or other canvas requirements.
- Processing Items: Only A-roll, or also captions, background music, transition sound effects, B-roll, MG animation, AI voice isolation, or export.
- Rhythm: Compact, natural, formal, relaxed, emotional, or follow the project's existing style.
Do not automatically include all processing items just because a tool is available. Leave enhancement items the user did not mention untouched.
When an existing project has canvas, caption style, or brand style, prioritize using them. Only use defaults when the project has no settings and the user has not specified requirements. Default for new projects is 1920×1080 / 60 fps; do not apply this default to existing projects, and do not change the aspect ratio on behalf of the user.
Landscape footage/canvas: During the final check and export phase (see section 12), remind the user if they need an additional portrait version. Portrait footage/canvas: Skip this, no need to ask every time. The judgment is based on the current timeline's canvas aspect ratio, not whether the user proactively mentioned it.
When multiple key dimensions are missing simultaneously, use a short form to collect them all at once; when only a small question is missing, ask a single core question directly. Do not repeatedly ask for information that can already be determined from the project or context.
2.1 Unified Storage of Editing Tables in Obsidian Script Document
Material tables (including B-roll checklist), sound effect tables, and transition tables generated or updated for the same video must be stored in the same Obsidian Markdown document as the video script, maintained in separate subsections; this document is the sole source of truth for these tables. Displaying tables in a chat does not replace saving.
- First, determine the video's script document from the user's Obsidian link, file path, or context.
obsidian://links need to resolve the vault and file, and locate the actual local Markdown file; if it cannot be uniquely determined, only ask for the target document, do not default to creating a new table file and do not write other videos into the known document. - Read the full content and existing subsections on the first access, and update them in place within the corresponding subsection; only append a new subsection if it does not exist. Keep the script, user annotations, and other content, avoid appending multiple copies of the table. Subsequent merges, deletions, renumbering, material matching, or timeline changes should all be synchronized to the same subsection and related summaries.
- The default columns for the material table are:
Paragraph / Voice Anchor | Required B-roll | Required Duration | Provided Yes/No | Corresponding Material. Paragraphs write the content topic and actual spoken sentences; duration is the number of seconds required for the final video, not the start/end time of the source video. Use ✅/❌ based on actual checked material, and note the check date and scope; do not claim absence without checking. When B-roll generation needs to be planned, add an extra columnSuggested Source (Pending Confirmation)in the total table to be confirmed, allowing the user to review both the visual needs and the source simultaneously; after confirmation, split into upload list and generation list. In other cases, do not add columns like "Can generate", "Visual change and landing", "How to connect". - The sound effect table records voice or action anchors, sound effect usage/type, duration, corresponding material, and confirmation/execution status; the transition table records the actual voice before and after the boundary, transition type, matching sound effect, and confirmation/execution status. Use the table structure the user has already confirmed. Only create tables when the corresponding processing items are enabled; saving a plan does not mean it is confirmed or placed on the timeline.
- Once the target document is specified, maintaining the above tables does not require asking for the save location every time. If old lists are in other locations, move them into the script document, leaving only a migration guide in the old file; do not maintain two sets of body text simultaneously.
- After saving, re-read to verify the table, number of items, numbering, duration, and status, and provide a document link. If writing fails, explain honestly; do not claim it was saved. Actual images, videos, and audio files are not embedded as binary in Markdown; record material paths or links in the table; material file location follows the user's storage convention.
3. Timeline State Machine
Phases are not mandatory; they are "enabled based on user selection." The only fixed dependency is that A-roll must precede all timeline-dependent processing items.
Recommended order:
- Project and material confirmation
- A-roll voice editing
- Captions
- Background music
- Transitions and matching sound effects
- B-roll
- MG animation
- Final check and export
If the user requests only one processing item, complete that item and perform related verification; do not automatically expand the task.
Phase Handover
After each enabled major phase is completed:
- Summarize the just-completed result in one sentence;
- Mark the current result as "pending user confirmation";
- Explain what the next step is;
- Ask only one core question that will affect the next step.
When the user says "continue", "next", "okay", or "confirmed", treat it as confirmation of the current pending phase and proceed to the next phase. If the user also requests modifications, complete the modifications and re-check the current phase first; do not skip the confirmation point.
When the user explicitly requests end-to-end completion, steps without subjective choices can be executed continuously; however, still stop for confirmation in the following situations: paid generation, irreversible operations, lack of key creative direction, or decisions that would significantly change the final video structure.
Do not bundle confirmation questions from multiple phases into a single message. Do not claim "completed" without actual execution or verification.
4. A-roll: Process Speech Content First
When judging whether a sentence should be kept or deleted, and how to present change proposals to the user, first consult references/speech-cleanup-judgment.md—it contains specific classification standards refined through repeated practical corrections (how to distinguish filler words/false starts/abandoned sentences/simplifiable content/evaluative remarks/operational steps), presentation format standards (table: left column original text with strikethrough, right column modified effect), and a common error checklist.
4.1 First Classify the Task
Before starting editing, determine what type the user wants:
- Clean: Remove fixed filler words, failed takes, speech errors, meaningless pauses, and obvious gaps.
- Refine: Extract key points, opinions, or highlights from long material.
- Reorganize: Adjust the order of opinions to make conclusions, conflicts, steps, or explanations clearer.
- Opening hook/short version: Bring strong opinions, results, conflicts, or questions forward, or compress to the target duration.
- Target script alignment: Retain and rearrange content according to the user's provided text, paragraphs, or specified sentences.
The tool cannot decide the editing strategy. First determine the content boundary, then choose the tool.
4.2 General Semantic Principles
- Process as much as possible by complete sentences, complete thoughts, complete answers, or complete steps.
- When the user explicitly specifies a sentence to keep, preserve the exact specified words precisely; do not carry irrelevant beginning or ending parts from the same ASR line.
- Do not splice fragments from different failed takes into an artificial sentence.
- Retain subjects, objects, conjunctions, transition words, list labels, qualifiers, and context necessary for a sentence to be valid.
- Repetition is not necessarily an error: it may be emphasis, structural marking, supplementary information, or emotional expression. Only failed attempts at the same intention should be treated as false starts.
- When unsure whether a cut point damages semantics or listening experience, keep the content or narrow the deletion range.
- When explaining edits to users, use actual spoken content or natural language; do not expose internal addresses like
[sN], word indices, clip IDs, segment IDs, etc.
4.3 Recommended Operation Order
- Use
read_scriptto read the current Script andtimeline.mdto first understand the content structure. - If there are fixed meaningless filler words or batch long pauses, first use
clean_scriptfor mechanical cleaning. - After cleaning, re-read the latest
timeline.md. Do not edit based on memory of the old transcript before cleaning. - Perform semantic editing: select the best take, delete false starts, remove duplicate or failed segments, and rearrange content if necessary.
- Use
apply_scriptto apply changes. - Re-read the generated transcript to check semantics, order, sentence breaks, repetition, and pauses.
- After the spoken-word structure is determined, run
smooth_audioto check and handle seams.
manage_transcript fix is only for correcting ASR text or speaker attribution; it does not cut audio. find_transcript is only for locating, not for editing. All deletions, moves, selections, and reuse of spoken content are done within the Script.
4.4 Selecting the Best from Multiple Recordings
Group recordings that express the same meaning, and compare group by group:
- Is the meaning complete and consistent with the target script?
- Fluency, swallowed words, stumbles, and tone.
- Is the breathing natural?
- Background noise, voice quality, visual stability, and cut point margin.
- After connecting with preceding and following sentences, does it sound like a natural single expression?
Preferentially keep one complete version per semantic unit. Do not assume the last recording is the best; only replace the previous version with a later one if it is more complete, smoother, or adds indispensable information. After applying, listen to cross-version seams one by one, focusing on stumbles, cut-off breaths, voice quality changes, and exposed noise.
4.5 Fillers, False Starts, and Half-Sentences
Pure hesitation sounds can usually be deleted in bulk: 呃, 额, um, uh, er, ah. The following words must be judged in context and cannot be mechanically deleted based on a list: 然后, 就是, 嗯, 那个, 所以, 但是, 啊, まあ, なんか, etc.
Keep a word when it serves as sequence, cause-effect, transition, reference, response, emphasis, or tone; delete it only when it is purely a filler and removing it still sounds natural.
For false start cleanup, follow this order:
- Confirm whether the several takes are indeed expressing the same intent.
- Find the most complete and natural version; do not mechanically choose the last version.
- Trim from where the failure or repetition actually begins; only cut off the covered or abandoned part.
- Retain the lead-in, conjunction, subject, comparison, or conclusion from the later version that is not repeated but makes the meaning valid.
For abandoned half-sentences like "This is actually...", delete them only if they do not provide useful information and would not make the following text abrupt. Do not delete necessary context just for brevity.
4.6 Pauses and Breathing
Compress pauses by default, do not zero them out:
- When not specified, focus on obvious pauses exceeding approximately
0.8–1 second, usually compressed to about0.3 seconds. - Keep about
0.3–0.5 secondsbetween sentences to allow listeners to perceive natural sentence breaks. - Topic shifts, comparisons, emphasis, and emotional transitions can retain longer pauses.
- Short breaths within sentences are usually kept; only significantly long pauses are compressed.
- Remove failed takes along with their preceding and following meaningless long pauses.
- Do not create unexpected
[gap]on the main video track just to create rhythm; it may render as black frames. When whitespace is needed, keep the source material muted, use[silence=...], or explicitly use B-roll, MG animation, or full-screen visual sections.
If the user provides explicit thresholds, use the user's thresholds.
5. Captions
Captions depend on the final A-roll. Do not submit final captions based on the old timeline when A-roll is not yet confirmed.
When there is no existing style and the canvas is 1080 landscape, use the following defaults:
- Style:
deyi-card; DeYiHei font, white text, dark semi-transparent rounded rectangle background card; - Position:
bottom-center; - Font size:
80 px, corresponding tofontSizeRatio: 0.074074; - Background card padding: horizontal
28 px, vertical18 px; - Rhythm: prefer short cards, naturally splitting by complete semantic groups;
- Scope: all audible spoken words on the timeline.
Adapt for other canvases proportionally; do not force 1080 landscape font sizes on them. Prioritize existing user preferences; do not ask repeatedly.
1080 landscape default is one line per caption card (maxLines: 1). If the text is too wide, split it into two or more short cards by complete semantic groups; do not wrap lines within a single card. The fixed priority for splitting is: Do not break words and set expressions > semantic completeness > single-line display > even card length. Rather than filling width, do not split fixed phrases, technical terms, English product names, numbers with units, percentages, or negative words from their predicates; avoid isolated 1-3 character caption cards unless for deliberate emphasis or speech timing constraints.
Before paginating captions, treat Chinese phrases, technical terms, product names, numbers with units, and fixed expressions as indivisible units. When a single line is too wide, move the entire group to the next card, falling back to the nearest punctuation, pause, subject-verb, verb-object, or parallel semantic boundary; do not hard-cut by character count.
After generating or re-splitting captions, perform a cross-card boundary review: for each card, check the last 2-4 characters of the previous card and the first 2-4 characters of the next card to identify split two-character words, multi-character words, English names, number-unit pairs, negative structures, and particle attachments. If a caption card ends or begins with an isolated character, or if the previous card ends with a highly dependent character like 的, 了, 和, 在, 不, 被, 把, review in context; when anomalies are found, fix by moving complete phrases or rebalancing adjacent cards; do not only process individual cases pointed out by the user.
By default, do not show Chinese comma ,, Chinese period 。, English comma ,, or English period . at the end of caption cards; retain question marks, exclamation points, decimal points, abbreviations, ellipses, and other punctuation within the sentence. The cleanup targets punctuation at the end of each caption card; do not delete commas or periods inside sentences used for sentence breaks or expressing tone.
After each generation, re-application, or modification of captions, and before export:
- Read all caption cards page by page; do not only look at the first batch.
- Use caption card text editing to actually save the end-of-card punctuation cleanup; do not replace it with a global toggle that hides all punctuation.
- Use the card IDs and latest revisions returned by the tool; do not guess batch interfaces.
- After saving, re-read page by page to confirm that the count of Chinese and English commas and periods at the end of cards is zero, and that other text and punctuation have not been mistakenly modified.
- Only if both saving and re-checking are successful, report the cleanup completion to the user, stating the actual number of modifications.
Pagination only modifies caption cards: use forcePageBreak:true for the first character of a new card, keepWithPrevious:true for merging into the previous card, and hidden:true for hiding repeated words or speech errors in captions. Do not fix a single cut point by modifying the original word-by-word transcript, font size, caption box width, or wordsPerPage.
After applying captions, send the complete captions to the user for manual confirmation, separated by | without sequence numbers; only after user confirmation proceed to background music or other downstream layers.
6. Background Music
Background music is optional. Based on the final spoken word, first identify the beginning, body, ending, and real emotional boundaries; do not default to using the same emotional template for all three sections.
Before generating or selecting a track, explain:
- The expressive function of each section, e.g., diagnosis, explanation, demonstration, reflection, humor, or call to action;
- Emotion and energy level;
- The actual content boundaries for music transitions.
First provide a small number of audition candidates with clearly different directions. Do not place on the timeline before the user selects a direction. After confirmation, produce the opening, body, and ending within the same sound system; transition points follow content boundaries, not equal segmentation.
When placing:
- Set the spoken-word track as
anchor, and the music track asfollower, using automatic ducking; - Unless explicitly requested by the user, do not manually set
audioRouting.duckDepthDb; - Do not significantly reduce the volume of music segments and then apply deep ducking;
- Fade in/out typically
1-2 seconds; - Music must not exceed the actual content ending;
- When music is insufficient, splice multiple segments on alternating tracks; do not unnaturally stretch a single segment;
- Do not place short sound effects on the follower music track.
7. Transitions, Transition Sound Effects, and Emotional Sound Effects
Enable only when the user needs transitions or sound effects. First distinguish the following concepts, then propose and execute:
- Transition effect: A visual change used to connect shots, distinguish semantic sub-points, or chapters. Ordinary straight cuts are not automatically considered transition effects requiring sound effects.
- Transition sound effect: A sound effect that always accompanies a transition. Every transition effect must have a corresponding transition sound effect; synchronize visual change and audio placement.
- Emotional sound effect: A sound effect that reinforces emotions or information with text, e.g., pain point, expectation gap, reminder, surprise, confirmation, achievement, benefit. It can appear independently of transitions and does not require a matching visual transition.
Sound effects are not necessarily tied to transitions, but when using a transition effect, a matching transition sound effect is required. Background music and audio cross-fades at spoken-word seams are not considered completed transition effects or the above two types of sound effects.
Choose based on content hierarchy and rhythm; do not overuse the same combination consecutively. For larger chapter switches, prioritize whip pans with clearly directional, slightly longer duration, accompanied by whoosh sounds, to signal to the audience that a new chapter has started; For closely related short steps or enumerated points (e.g., first, second, third step), use short mouse clicks and light visual beats, emphasizing they still belong to the same content group. Duration should reflect priority but not have a fixed number of seconds universal to the whole video; first listen to the spoken word and look at the visuals, then determine specific length and cut point.
- Quick flash to black + mouse click sound effect: Suitable for tightly related step transitions; the main transient of the click aligns with the most obvious visual node of the flash. The flash to black must be short; do not make obvious chapter breaks between steps.
- Whip pan + whoosh sound effect: Suitable for transitioning into larger chapters; the horizontal motion blur should be clear, the duration significantly longer than step clicks, the whoosh's motion direction aligns with the visual, and the peak or landing point aligns with the key node. Do not mistake a linear wipe with a white dividing line for a whip pan or black layer; if a jarring white line appears in preview, replace it with a real whip pan or the user-specified fade transition.
Listen and preview after placement. Transitions must be perceptible but not dragging; visual changes and audio landing points must be synchronized; sound effects must not cover the spoken word.
7.1 Transition Points: Proposed by the Agent Each Time, Confirmed by the User, No Fixed Rules
Different video scripts have different structures; transition points cannot be identified by mechanical rules like pauses, blank lines, or keyword matching. The essence of a transition point is a switch between semantic sub-points within the same content (e.g., where enumerated points like "first/second/and another" in a list each begin, or where a complete example ends and the next feature point begins). This must be judged by understanding the content semantics, not by structural proxies.
Before placing each transition:
- Read the entire current Script/transcript, find candidate positions that are semantically suitable for transitions.
- Organize into a plan and send to the user for confirmation: quote the text before and after candidate points using actual spoken content (do not show frame numbers or internal addresses like
[sN]), explain the suggested transition effect and corresponding sound effect, and present the transition table separately according to section 7.3. When emotional sound effects are needed, list an emotional sound effect table separately. - After user confirmation or adjustment, place the confirmed transitions and sound effects on the timeline; do not place without confirmation. "Execute according to these two tables" by the user authorizes execution of the latest version of both tables; do not ask again about confirmed points.
- Once a plan is confirmed, it becomes the transition baseline for this video. Subsequent adjustments for the same video (including landscape-to-portrait conversions, adding opening hooks, etc., as long as the body content hasn't changed) follow this baseline; do not re-propose the entire plan each time.
7.2 After User Adds or Deletes Content, Only Check the Affected Parts
When the user subsequently adds or deletes spoken content, do not restart a full round of transition plan confirmation; instead:
- Re-read the current Script, check one by one whether the text anchoring each confirmed transition point still exists and whether the semantic boundary is still valid.
- If the anchoring text still exists and only the frame number has changed: directly move the transition to the new frame number; do not ask the user again.
- If the anchoring text has been deleted or rewritten, causing the original boundary to no longer be valid: only tell the user about these 1-2 affected points, specifically which text changed, and ask whether to change to an adjacent anchor or delete this transition—do not involve other unaffected transition points.
Derivative versions like landscape-to-portrait and adding opening hooks also apply point 2: as long as the body content has not changed, directly re-locate to the corresponding frame numbers on the new timeline based on the confirmed baseline; no need for a full re-confirmation.
7.3 Two Separate Execution Tables, Not Merged with B-roll Material Table
By default, maintain separate B-roll material table, transition table, and emotional sound effect table. They can reference the same spoken-word anchor, but having a B-roll change at a certain point does not automatically add a transition or emotional sound effect.
Transition Table: Each row must simultaneously contain the transition effect and the corresponding transition sound effect.
| Approx. Time | Voice Transition Point (Before → After) | Transition Effect & Duration | Corresponding Transition Sound Effect | A/V Sync Method | Status |
|---|---|---|---|---|---|
| Based on current timeline location | Using actual spoken word quotes | Flash to black / Slide right, etc. | Click / Whoosh, etc. | Transient, peak, or landing aligned visual node | Pending / Confirmed / Added / Verified |
Emotional Sound Effect Table: Based on the meaning of the text and emotional landing points; can exist without any transitions.
| Approx. Time or Range | Text Landing Point | Reinforced Emotion or Information | Suggested Emotional Sound Effect | Usage | Status |
|---|---|---|---|---|---|
| Based on current timeline location | Precise keyword or short phrase | Pain point / Reminder / Surprise / Sense of completion, etc. | Description of sound characteristics | Number of occurrences, landing point, tail sound, and volume requirements | Pending / Confirmed / Added / Verified |
The time in the table is for lookup; actual execution is based on the latest spoken-word anchor. First fully understand the semantics, then align word by word in time; do not automatically decide which words must have sound effects based on keywords. Do not solidify the current project's time points, quantities, or material IDs as default rules for other videos.
7.4 Execution and Verification
- Use separate, clearly named audio tracks for transition sound effects and emotional sound effects; do not mix them into the background music follower track, allowing independent adjustment.
- Emotional sound effects should be short, light, and few; choose the most important one within the same semantic unit to avoid stacking across every word. Check for overlap between adjacent transition sound effects and emotional sound effects; stagger main transients or shorten tail sounds; do not cover the spoken word.
- The "acceleration priority" rule for screen recordings does not require long tail sounds for sound effects; sound effects can be shortened by listening perception and given a natural fade-out.
- Transitions act on the visuals the audience actually sees. Check the B-roll, base layer, and other tracks above; do not claim the transition is visible if you only add an effect to the covered spoken-word track. Maintain the timing of spoken word, captions, and other content.
- When the user cancels a transition point, cancel the transition effect and the corresponding transition sound effect in pairs; emotional sound effects remain as independent suggestions unless the user requests adjustment.
- After execution, cross-check the latest points, quantities, pairings, and track distribution in both tables; preview entry, key nodes, end, and adjacent visuals one by one, while playing to check A/V sync, volume, and tail sounds. If not actually listened to, clearly state this; do not treat position checking as audio verification.
- Report the quantity of transition effects, transition sound effects, and emotional sound effects separately; distinguish "added" from "verified"; do not report only music or spoken-word cross-fades as transitions and sound effects completed.
8. B-roll
B-roll is used to cover natural jump cuts, show specific objects, or enhance understanding. First confirm the mode:
Table first, production later: When the user plans to plan or generate a batch of B-roll, first read the current official spoken word, existing storyboards/scripts, and actual material library; organize a unified visual material table, listing spoken-word anchors, required visuals, suggested display duration in the final video, whether usable material already exists, corresponding material, and suggested sources. Save it in the same Obsidian script document, clearly marked as pending confirmation; the suggested source is just a proposal, not authorization for generation. After the user confirms or adjusts the visual needs and sources, split the confirmed version into two execution lists: real material the user needs to upload (websites, tool operations, personal works, documents, covers, etc.) and material the Agent can generate (concept scenes, cartoon supplementary visuals, etc.). Real product operations or results cannot be faked with AI concept visuals; items the user is to upload cannot be changed to generation items. Then, respectively collect/verify uploaded material, or proceed with generation material based on the generation skill and necessary paid confirmation, followed by placement and acceptance. Do not mistake list confirmation for authorization for generation, download, replacement of real evidence, or timeline modification.
- Full-screen switch: Suitable for covering jump cuts, displaying content that requires complete reading or viewing.
- PiP small window: Retains the talking head while providing auxiliary visuals.
- Agent decision: When the user has no preference, decide per content point.
Before placing PiP, check the target frame and material frame, avoiding faces, mouths, key gestures, captions, products, logos, and readable text; default rounded corners 24–36. For full-screen material, choose cover or contain based on aspect ratio and protected content. By default, do not cut away within the first and last 3 seconds unless content requires it.
Sources can be project material, stock material, or AI video. When paid generation is involved, obtain explicit user authorization before submitting for generation. After placing each B-roll, check the visuals to confirm that the subject, captions, or key operations are not obstructed.
8.1 Material Table-Driven B-roll
When the user provides a B-roll material list, checklist, storyboard, or a table of "spoken-word anchor—required visuals—suggested duration", and requests placing material library content at corresponding positions, use references/b-roll-checklist-workflow.md.
The core rule for this type of task is: first verify actual footage and match it according to spoken-word semantics, then decide on speed and coverage range; prioritize keeping the video fully and accelerating it, only editing if it still cannot fit within the related spoken-word interval within the allowed speed range. Do not simply cut a short segment from a long video for convenience.
8.2 Text Screenshots: Enlarge the Body Text and Scrolling Reading
- For screenshots primarily consisting of text, such as chat logs, user answers, reports, etc., prioritize enlarging the body text related to the spoken word rather than displaying the entire interface at a reduced size. First check actual pixels, determine the body text boundaries, and then crop out irrelevant sidebars, toolbars, and blanks; do not cut off the title, source, context, or key text needed for understanding.
- When the body text is long, use screenshot panning to simulate top-to-bottom browsing: move the visual content upward, revealing the text below sequentially; do not move text downward, and do not describe a static screenshot as a real operation screen recording.
- Brief pauses at the beginning and end to allow viewers to orient and read; arrange pause, scroll distance, and speed according to the length of the body text, information focus, and current spoken-word coverage duration; do not use a fixed number of seconds. If a short interval cannot display clearly, narrow the scroll range, only display relevant content, or suggest adjusting the duration; do not scroll too fast just to scan the whole page.
- In principle, maintain the timing of confirmed spoken word, captions, music, and adjacent material; use editable position keyframes or equivalent deterministic animation. During the entire movement, do not show black edges, and do not crop horizontal text; check for caption obstruction of key text.
- Check the entry, middle of scroll, end, and adjacent cut points, and continuously preview the scroll speed when playback is possible. If only checked at irregular intervals, explain truthfully; do not treat frame sampling as motion audio verification. Synchronize actual composition, display range, and duration with the material table in the same Obsidian script document.
- This rule applies to text screenshots; do not automatically crop or repackage already composited screen recording B-roll.
9. MG Animation
Consistent style within the same video (default, no user emphasis needed): All MG animations for the same video (including prompt pages, post highlights, titles, tags, and chapter progress bars) must adhere to the confirmed unified design style. Before production, check the existing MG for this video, maintaining consistency in color scheme, font, font weight hierarchy, corner radius language, lines, and animation rhythm; compare multiple proposals only within the same visual system structure. Do not import inconsistency by reusing templates from other videos; do not solidify this video's specific colors as default colors for other videos. Pass this set of visual constraints to xb-video-timeline during chapter progress bar handover.
MG animation is an optional visual interpretation layer and should be done after determining A-roll, captions, and the direction for music and B-roll as needed. First determine what it helps the audience understand based on the script and target frames, then decide:
- Content: title, tag, list, comparison, diagram, or full-screen information;
- Timing: sync with which spoken beat or information point;
- Form: transparent overlay or opaque full-screen;
- Position: avoid faces, mouths, key gestures, captions, and existing elements.
When creating MG animation, use the Motion Graphics workflow available in the current ChatCut environment. After placement, check the stable frame; do not mistake temporary absence in transition animation for a design error. Full-screen MG can intentionally replace A-roll, but this must be a clear content decision.
MG Production Workflow: First read the latest transcript and target visuals; decide whether to use full-screen MG or transparent overlay; follow the project's existing design style; first produce an actual viewable HTML static draft; after user confirmation of the static draft, generate the animation, place it on the timeline, and check keyframes. Do not directly enter animation production when the static direction is not yet confirmed.
Multi-proposal review: When the visual direction is not yet determined, first provide 3 clearly different HTML static proposals for the user to review; when the direction is already confirmed, iterate only around that proposal; do not diverge into other directions.
Composition Rules: MG avoids captions, faces, and key operation areas; when captions are present at the bottom, shift the visual center of the MG main content up by approximately 40-60 pixels to leave clear space for captions.
MG to Video Specifications: Whenever exporting or baking MG animation to video, use 1080p, 60 fps; after export, directly replace the corresponding MG on the timeline's B-roll track and perform real export frame sampling acceptance.
MG Acceptance: Check three states: entry, stable, exit; for chapter-type MG, also check each chapter transition boundary. During acceptance, simultaneously verify caption avoidance, background coverage completeness, and whether the MG maintains the correct aspect ratio on the current canvas without being compressed or deformed.
9.1 Chapter Progress Bar
The chapter progress bar is no longer fully designed and animated within this skill. When a top/bottom chapter timeline, chapter navigation, or progress indicator that advances with playback is needed, hand over to the independent skill xb-video-timeline to avoid xb-video-cut simultaneously maintaining video cards, chapter semantics, and progress bar visuals, causing rule duplication and upstream/downstream inconsistency.
Before handover, this skill organizes and confirms:
- Chapter names, order, and final copy;
- Real start and end positions of each chapter in the final A-roll timeline;
- Actual content end points, not the total timeline length formed by music or caption tails;
- Final video aspect ratio, caption position, and avoidance requirements for faces and key operations;
- Confirmed Design Style, primary colors, and reusable reference styles.
After handover, xb-video-timeline is responsible for HTML style review, progress visuals, animation generation, time mapping, timeline placement, and chapter boundary acceptance. This skill does not redundantly generate progress bar proposals or copy its visual parameters and implementation details.
If A-roll, chapter titles, or video cards change subsequently, this skill only updates the chapter content and boundary list, and passes the latest result back to xb-video-timeline; do not retain old frame positions, and do not maintain two sets of chapter data across both skills.
9.2 Chapter Titles: Title First, Then B-roll
When chapter titles are enabled, by default display the title in an existing blank area of the A-roll spoken-word visuals, and only cut to the corresponding chapter's B-roll after the title has completely disappeared. The title's exit animation must also finish; do not let a large title compete for reading attention simultaneously with body text screenshots or operation interfaces.
- Maintain the original A-roll person's position, scale, and crop; adapt the title to existing blank space; do not move the person for the title. Avoid faces, key gestures, captions, and operation areas.
- When arranging the start time for subsequent B-roll, simultaneously check the base layer and other related layers; do not only move the image later while letting the black base layer block the person prematurely. Keep the confirmed spoken-word timing; review material coverage and adjacent cut points; do not arbitrarily cut off real screen recording operations to free up time for the title.
- Confirm title text, chapter count, and landing points from this officially spoken word; do not reuse chapter counts, numbers, or time points from other videos; do not automatically replace confirmed transitions and sound effects by adding titles.
- Use the previous static draft confirmation process: first overlay representative titles on the actual spoken-word visuals for inspection before batch application. Only create opening titles, black chapter cards, or permanently visible navigation when the user explicitly requests it.
Confirmed Reusable Title Style: Pale Yellow Number / Off-White Pixel Font
Used when the user selects this pixel chapter title direction, or when enabling chapter titles without another confirmed title style; existing project design style and this specific request take precedence; do not overwrite all MG with this scheme.
| Element | 1920×1080 Reference Specification |
|---|---|
| Font | GuanZhi-8px pixel font; local actual family GuanZhi 8px, file <local font directory>/GuanZhi-8px.ttf |
| Chapter Number | Two-digit number, e.g., 01; approx. 120px; pale yellow #F6DFA0 |
| Chinese Title | Approx. 80px; off-white #F4EEE4 |
| English Chapter Description | Approx. 26px; off-white #F4EEE4; corresponds to the chapter; do not fabricate descriptive facts; if space is insufficient, prioritize omitting it |
| Weight & Edge | Regular weight, no faux bold; no drop shadow, blur, or stroke; keep pixel edges clear |
| Composition | Transparent overlay, number, Chinese, English left-aligned; prioritize existing blank space next to the person, approx. 6% from the left; no unrelated decorations or default background cards |
| Appearance Order | Number appears first, then Chinese title, then English; no trailing vertical cursor after reveal completes |
The above font sizes are starting references for 1080 landscape; adapt to other canvases by actual display ratio; when necessary, shrink or wrap based on available blank space, do not stretch glyphs. Do not check only the standalone design draft; must accept font size, spacing, readability, and avoidance using actual spoken-word visuals. Report missing fonts; do not replace silently; confirm the actual rendered font is the specified one.
This subsection records the title font sizes, color scheme, and no drop-shadow requirement confirmed on 2026-11-07; in case of differences with old xb-video-style title values, use the new specification in this subsection, but do not change other video layers, transitions, or caption styles as a result.
10. AI Voice Isolation and Audio Seams
When cleaning or isolating vocals is needed, use AI voice isolation to process the vocal track, keeping the video visuals unchanged. First confirm the user indeed wants to process audio; do not treat AI voice isolation as a default step.
After completing A-roll structure, music placement, or other timeline-altering operations, check spoken-word seams. After the user manually deletes clips, re-read the tracks, only close accidental gaps, then run smooth_audio. Do not use fades to mask wrong version choices, wrong cut points, or tone changes.
11. Final Verification and Delivery
Before export, check one by one based on the enabled processing items:
- No accidental gaps, black frames, or content duplication;
- Spoken-word seams are natural, semantics complete, pauses not too tight;
- Captions sync with final spoken word, end-of-card punctuation review passed;
- Background music does not exceed the content ending, vocals clear, ducking normal;
- B-roll and MG animation do not obstruct the subject, mouth, gestures, captions, logos, or key text;
- Transitions and matching sound effects are synchronized and do not cover the spoken word;
smooth_audiohas been re-checked after the last structural change;- Canvas, resolution, frame rate, and export specifications match the project goal.
11.1 Before Export: Archive Generated Materials and Re-link
Before exporting the final video, add a step to centrally archive the materials generated during production of this video to the corresponding folder under the user-specified material archive root directory, and have the project reference the archived files:
- Determine folder: Based on the source spoken-word material folder for this video under
<Material Archive Root Directory>/<Video Folder>/(e.g.,example-video/). If it cannot be uniquely determined, only ask for the target folder; do not create a new location. - Archive scope: AI-generated images, keyframes, and videos, cutout layers and filled backgrounds, generated voiceovers, music, and sound effects, MG converted to video, and all other generated materials. Create subfolders by shot or purpose (e.g.,
S004-cartoon-scene/still-frame,S004-cartoon-scene/cutout-layer), use readable names from the material library, do not retain hash-style filenames. User-uploaded original materials and temporary inspection screenshots are not included. - Copy, do not move: Copy from the original location such as tool cache, project working directory; do not delete the original files.
- Re-link: Re-link references in the material library one by one
...
[truncated for translation]