Knowledge Video Editor · Intelligent Talking-Head Video Editor
Make knowledge visible, logic clearer, and viewing less tedious.
This Skill transforms "an unedited talking-head video" into "a publish-ready knowledge video." You just provide the video, and it autonomously performs: understand content → determine type → fine-cut audio → direct and plan → visualize logic → add bilingual subtitles → generate first-frame cover → quality check → deliver final cut.
Core principle: Fully automatic by default, no step-by-step confirmation required; but never fake the final cut.
1. Determining Whether to Use This Skill
| Use It | Don't Use It |
|---|---|
| Real or digital presenter talking-head videos (explaining knowledge, methods, or opinions to camera) | Pure vlogs, travel clips (no explanatory narrative) |
| Existing talking-head footage needing repackaging as a knowledge video | Pure film editing, narrative films, short dramas |
| Need horizontal/vertical final cut, bilingual subtitles, and cover frame | Only need transcoding, cropping, or resolution change |
| Batch standardized output (same visual style) | Need to keep original and only add subtitles (use basic subtitle workflow) |
Judgment basis: Does the video contain "information that can be presented in a structured way"? If yes, proceed with this Skill.
2. Three Golden Rules Before Starting
- Check the environment first, then act. Execute
python3 scripts/check_environment.py --jsonto confirm ffmpeg / transcription engine / rendering engine / fonts are ready. If critical capabilities are missing, report the actual gaps—never "pretend" to produce a final cut. - Plan first, then render. Must generate
director_plan.jsonandstoryboard.json(per-second scene breakdown), then validate withvalidate_storyboard.pybefore rendering. - Leave original footage untouched. All intermediate artifacts go under
<project>/work/, final output under<project>/output/. Never overwrite the user's original video.
3. 13-Step Execution Pipeline
| Step | Action | Output | Script |
|---|---|---|---|
| 1 | Environment self-check | env_report.json | check_environment.py |
| 2 | Video property detection (orientation/duration/resolution/framerate/presenter position) | video_meta.json | analyze_video.py |
| 3 | Content understanding and segmentation (determine overall type + per-segment attributes) | analysis.json | analyze_video.py --content or host LLM |
| 4 | Speech transcription + editing decisions (repetitions/verbal errors/pauses/filler words/beginning-end) | transcript.json + edl.json | transcribe.py |
| 5 | Fine-cut execution (trimming + noise reduction + loudness normalization + timeline remapping) | edited.mp4 + timemap.json | edit_video.py |
| 6 | Full video director plan (video type/topic/cover/chapters/pacing) | director_plan.json | LLM + templates |
| 7 | Logic visualization (knowledge points → visual component mapping) | visual_map.json | LLM + 07 specification |
| 8 | Per-second storyboard (every 1 second planned, composition reviewed every 3 seconds) | storyboard.json | LLM + storyboard.schema.json |
| 9 | Adaptive layout (horizontal/vertical safe areas, presenter avatar region) | layout.json | config/safe-areas.json |
| 10 | Chinese-English bilingual subtitles (sentence-level, semantic breaks, width measurement) | subtitle.srt + subtitle.json | transcribe.py --bilingual |
| 11 | First-frame cover generation (first frame is the cover) | cover.png | render_video.py --cover-only |
| 12 | Final rendering (Remotion composition + presenter avatar + sound effects/BGM mixing) | output/01_FinalCut.mp4 | render_video.py |
| 13 | Quality check and auto-fix (content/visuals/subtitles/audio/decoding) | qc_report.md | quality_check.py |
One-click pipeline: python3 scripts/run.py --video <source> --out <directory> --ratio 16:9
4. Non-Negotiable Hard Rules (Violation = Rework)
- The first frame must be a complete cover. No black screen, blank space, or title still in entrance animation. Title max 2 lines, each line max 4 characters (letters/digits counted same), total ≤ 8 characters, centered, extra-bold large font.
- From the second shot, the cover title must completely disappear. Cover defaults to 1.5–2 seconds, synchronized with speech start—no silent waiting inserted.
- Horizontal must include dynamic chapter navigation bar. 3–6 chapters, each 2–6 characters, smoothly switches with speech progress, max 8% of frame height, hidden during cover. Navigation bar uses capsule floating style (border-radius = height/2, frosted glass, 1px stroke, sliding highlight block, bottom thin progress line). No "text+checkmark+dash" stacking allowed.
- Whole video uses only one main color scheme. Deep Space Black Gold or Pearl Ink Green, no mixing.
- Bilingual subtitles: Chinese above, English below, sentence-synced, must measure character width—no overflow, no blocking core knowledge area, no infinite font size reduction.
- Audio priority: original voice > key sound effects > BGM. BGM must be low volume and duck behind voice.
- Don't delete key arguments to shorten duration, don't change original opinions, don't rewrite facts.
- Don't crop horizontal output directly to vertical—must re-layout and re-render.
- Quality score < 90 must auto-locate and fix; if unfixable, report honestly—never declare pass.
- All assets must be licensable—no images/music/video from unknown sources.
- Every frame must include the presenter. Presenter shots use V01/V02 full frame; knowledge visuals (keywords, flows, comparisons, cards, numbers, summaries, etc.) must overlay V21 circular presenter avatar in the top-right corner—no frame allowed without presenter.
- Three iron rules for circular avatar framing and position. (a) The circle must show hair + full facial features + shoulders, with 6–10% headroom above and at least 25% space below chin for neck and collar—no forehead or chin cropping; (b) Fixed top-right, vertical below navigation bar, horizontal at top-right edge, with 48–52px margin; (c) Must never overlap title, knowledge cards, or subtitles—content area must yield to it (vertical content shifts down, horizontal right side leaves 344px safe zone). Three framing parameters:
zoom(inner width / circle diameter, 1.0 standard for head-shoulders, use 1.2–1.5 only if person appears small) +focus_x/focus_y(face center) +anchor_y(face center vertical position in circle, close-up 0.42 / wide shot 0.28); must render one still frame for acceptance before final. - Explanatory text defaults to 12% larger. Knowledge visual titles, bullet points, card body rendered uniformly at
text_scale 1.12—prefer reducing words over shrinking font.
5. Video Type Determination (Not User-Selected, But Determined by You)
Judge overall video first, then segments. Different strategies can be mixed within the same video.
| Type | Preferred Editing Style | Presenter Ratio Reference |
|---|---|---|
knowledge Science popularization | Flowcharts, structure diagrams, dynamic whiteboard | 20–35% |
practice Tool tutorial | Screen recording, local zoom, operation demo | 15–30% |
analysis Business analysis | Data charts, comparison charts, causal chains | 25–40% |
opinion Opinion sharing | Presenter + keywords + cases | 40–60% |
story Storytelling | Presenter narrative + scenes + timeline | 45–65% |
industry Industry interpretation | Maps, citations, real footage, charts | 25–45% |
Ratio is a reference, not a check-in. Wrong type judgment = wrong overall style; this step determines 80% of the work.
6. Directory Structure
knowledge-video-editor/
├── SKILL.md # This file: single entry point, rule master
├── AGENTS.md # Execution protocol for each Agent (Codex/Claude/Gemini/Doubao, etc.)
├── README.md # Human-readable instructions and showcase
├── VERSION
├── references/ # 10 detailed rule documents (no code)
├── config/ # Machine-readable config (theme/ratio/safe zones/glossary/schemas)
├── scripts/ # 8 executable scripts (analysis→editing→validation→rendering→QC)
├── templates/remotion/ # Visual rendering engine (React components)
├── assets/ # Fonts/icons/music/sound effects (licensed assets, leave empty in repo with note)
└── examples/ # 4 real director plan examples
7. Reference Index (Read on demand—don't ingest all at once)
| Need to Solve | Read |
|---|---|
| How to determine type and segment | references/01_VideoUnderstandingAndTypeJudgment.md |
| What to cut, keep, and how not to cut wrong | references/02_AutomaticFineEditing.md |
| How to make cover that won't get rejected | references/03_FirstFrameCoverSpecification.md |
| Horizontal/vertical layout and navigation bar parameters | references/04_HorizontalVerticalAndDynamicNavigation.md |
| Two theme color values and usage scenarios | references/05_ColorAndVisualDesign.md |
| Per-second storyboard writing and animation curves | references/06_IntelligentStoryboardAndPerSecondDynamics.md |
| How to choose V01–V20 components, how to make V21 avatar | references/07_VisualComponentUsageSpecification.md |
| Subtitle font size, line breaks, safe zones | references/08_BilingualSubtitlesAndSafeZones.md |
| BGM and sound effect density, ducking | references/09_BackgroundMusicAndSoundEffects.md |
| How to score QC 100 and how to fix | references/10_QualityCheckAndAutoRepair.md |
8. Command-Line Quick Reference
# 1. Environment check
python3 scripts/check_environment.py --json
# 2. Read video properties
python3 scripts/analyze_video.py --video input.mp4 --out work/video_meta.json
# 3. Transcribe (Chinese-English bilingual)
python3 scripts/transcribe.py --video work/edited.mp4 --bilingual --out work/
# 4. Fine-cut per EDL
python3 scripts/edit_video.py --video input.mp4 --edl work/edl.json --out work/edited.mp4
# 5. Validate storyboard (no pass, no render)
python3 scripts/validate_storyboard.py --storyboard work/storyboard.json
# 6. Render (including cover export, circular presenter avatar)
python3 scripts/render_video.py --storyboard work/storyboard.json --ratio 16:9 \
--audio work/edited.mp4 --cover cover.png --out output/01_FinalCut.mp4
# 7. Quality check
python3 scripts/quality_check.py --video output/01_FinalCut.mp4 \
--storyboard work/storyboard.json --report output/08_AutoQCReport.md
# 8. Full pipeline
python3 scripts/run.py --video input.mp4 --out project/ --ratio 9:16 --theme pearl-green
All scripts support --dry-run to preview intended actions.
9. Data Structures (Must Follow Contract)
Director plan and storyboard use a unified JSON contract: config/storyboard.schema.json. Minimal usable structure:
{
"video_type": "knowledge",
"aspect_ratio": "16:9",
"resolution": [1920, 1080],
"theme": "dark_gold",
"cover": {
"lines": ["工作方式", "彻底变了"],
"position": "center",
"visible_from_first_frame": true,
"duration": 2
},
"presenter": {
"always_visible": true,
"badge": { "enabled": true, "diameter_px": 192, "top_px": 316, "focus_x": 0.58, "focus_y": 0.30, "zoom": 1.0, "anchor_y": 0.42 }
},
"navigation": { "enabled": true, "chapters": ["现状", "逻辑", "案例", "总结"] },
"subtitles": { "languages": ["zh", "en"], "layout": "chinese_above_english", "safe_area_check": true },
"storyboard": [
{ "scene_id": "S001", "start": 0, "end": 2, "visual_type": "cover",
"motion": "subtle_push", "sfx": "soft_impact" },
{ "scene_id": "S002", "start": 2, "end": 5, "visual_type": "presenter",
"cover_visible": false,
"animation_plan": [
{ "time": 2, "action": "show_navigation" },
{ "time": 3, "action": "highlight_keyword" },
{ "time": 4, "action": "subtle_zoom" }
] }
]
}
During actual execution, must also include: asset absolute paths, exact frame numbers, real subtitle timestamps, graphic coordinates, animation curves, licensing info, horizontal/vertical safe zones, and avatar framing parameters.
10. Cross-Agent Adaptation
The same rules are readable by any Agent; whether final cut can be produced depends on the environment.
| Agent | How to Read | Can Render |
|---|---|---|
| Codex / Claude Code / Gemini CLI | AGENTS.md → SKILL.md | Yes (requires ffmpeg + node) |
| Doubao Work / WorkBuddy | This SKILL.md | Depends on environment |
| Text-only LLM | Output only director_plan.json + storyboard.json | No, pass to rendering environment |
Division of labor: reasoning belongs to the model, rendering belongs to the program. Don't let the LLM guess pixels, and don't let the rendering engine make semantic judgments.
11. Delivery Checklist
<topic>/
├── 01_FinalCut.mp4
├── 02_FirstFrameCover.png
├── 03_FineCutAudioVideo.mp4
├── 04_ChineseEnglishSubtitles.srt
├── 05_FullTranscription.md
├── 06_DirectorPlan.json
├── 07_Storyboard.json
├── 08_AutoQCReport.md
└── 09_EditableVideoProject/
When outputting both horizontal and vertical, each gets its own final cut, cover, storyboard layout, and subtitle project, sharing content analysis and fine-cut timeline.
12. When Must Stop and Report
- Transcription engine unavailable with no alternative → report, don't make up lines.
- Source video severely blurry/without audio/without valid explainer content → report, don't force-edit.
- Asset copyright unclear → don't use, and explain why.
- Rendering fails multiple times → report the error verbatim, don't hand over a half-baked product as final.
Honesty matters more than completion.