Vane — Faceless Knowledge Short Video Skill
Boundaries
This process is executed by a single person with an AI capable of file handling and video production. It does not rely on a fixed team, dedicated platform, or specified paid tools. This is a production workflow with human confirmation nodes, not a product that guarantees auto-publishing from a topic. Without confirmation, no step proceeds. AI with only text chat capabilities can complete the planning phase but must not pretend files have been generated. The public version defaults to faceless: voiceover + real screen recordings/screenshots + journal cards + generic graphics. Live-action video or personal stickers are optional branches and require the user to provide and license them.
0. Confirm Task and Directories
Read README.md, docs/workflow.md, docs/visual-style.md, and docs/materials.md. Clarify the topic, audience, publishing platform, aspect ratio, duration, available materials, whether a script and voiceover exist, whether rewriting is allowed, and the deadline. Default for knowledge-based short videos: landscape 1920×1080, 30fps; 3:4 is 1080×1440; 9:16 is 1080×1920. Platform does not equal fixed ratio; follow this request's requirements. Create directories: project/raw, voice, assets, work, out; all temporary files go in the project work folder, do not use shared temp directories, and do not modify original materials. Save confirmed scripts word-for-word; do not rewrite, re-record, or add call-to-action phrases without permission.
1. Topic Selection, Fact-Checking, and Titles
Choose a specific question and clarify what the audience will take away. Prioritize real cases, tested processes, before-and-after comparisons, and reproducible steps. Record sources, access dates, applicability conditions, and original location for facts; numbers and quotes must be traceable. If uncertain, delete, mark, or verify with the user. Do not write someone else's test as your own experience; do not fabricate group chats, backgrounds, bills, success rates, or demo results. Draft 3 publishing titles with cover main and sub-headlines. The user selects one; confirm 1–2 actionable points worth saving. Avoid marketing clichés; never claim any word is certain to limit traffic or guarantee safety; verify current platform rules before publishing.
2. Script Lock and Voiceover Confirmation [Checkpoint 1]
Show the result in the first 3 seconds; give the first practical point within 15 seconds. Keep total duration under 3 minutes; for over 2 minutes, plan chapter sections. Script structure: Result → Specific Problem → Steps/Comparison → Limitations → Conclusion. Do not distort facts for a hook. Prioritize the speaker's own tone for live recordings; TTS can use any legal service or local tool; do not copy unauthorized voices. Segment according to the service's maximum character limit. Export the original-speed version first, trim seams and meaningless long silences, then apply speed changes as confirmed. This style's voiceover defaults to 1.2x speed with unchanged pitch for preview; do not speed up already recorded human voiceovers without authorization. Proofread English proper nouns and numbers during the full trial listen. Deliver the complete voiceover and text first. The user must approve before proceeding to material production; revising the text requires synchronized updates to voiceover and subtitles.
3. Material List Confirmation [Checkpoint 2]
Based on the approved voiceover, create a sentence-by-sentence or paragraph-by-paragraph material list; do not create a full storyboard first. Use templates/materials.csv: distinguish between ready-made materials, screen recordings/screenshots needed, cards to be newly drawn, and materials to be provided by the user. For each item, specify the corresponding line, purpose, source, license, dimensions, transparent background, needed layers, availability status, and responsible person (can be yourself for all). Prepare paper backgrounds, hand-drawn window frames, step bars, thematic corner markers, comparison/list cards, necessary icons, subtitle styles, legal music, and a few sound effects. The faceless version does not require photos or personal videos; optional generic little people may be used. Live appearance requires separate original video; stickers require properly licensed transparent cutouts. Blur or mask identities, account numbers, balances, and other unnecessary private data in screenshots. Charts without real data must be clearly marked as "illustration"—do not pretend they are backend screenshots. Once the user confirms the list, proceed to acquire or create materials. If materials are missing, list the gaps; do not silently substitute with unrelated visuals.
4. Storyboard After Materials Are Ready [Checkpoint 3]
Fill in templates/storyboard.csv according to the final voiceover timeline, listing main visuals, transitions, actions, subtitles, sound effects, and sources. Provide a storyboard overview for quick user review. Do not proceed to this step without confirmed script and material list. Visuals should change with the information but do not add effects for every word. Real contrast or an action result can be shown every 15–30 seconds to avoid empty segments. Center screen recordings inside hand-drawn window frames; show the current step on the step bar; highlight one key point at a time.
5. First 3-second Sample [Checkpoint 4]
First create the first 3 seconds using the final canvas, fonts, and voiceover. Check rhythm, readability, and whether the opening frame and audio convey the same message. Only proceed to the full video after approval to avoid rework. Sample renderer supports --seconds 3; render the full segment only after the sample passes.
6. Full Video Editing
Use Remotion, Manim, Pillow + ffmpeg, or a familiar editor—choose one method that works; do not force a specific tool. Reference docs/visual-style.md: cream white, charcoal black, orange-red, navy blue; minimal background decoration. White or content-matching color accent lines; do not overlay text. Layered elements: fade in and out, 0.3–0.4 sec bounce, ~8% overshoot as a starting point; stagger layers in the same group by 2–3 frames. Minimize screen shake; avoid PPT-style hard cuts with stacked images. Only add sound effects on the first entry of the main card and key text hits—not every layer. Background music approximately 20dB below the voice as an initial listening setting. For face cutouts: directly remove the background, do not frame them. Only use picture-in-picture frames for real video; fill the frame, show the upper body, do not let the head fill it; emphasis push does not exceed 110%. Do not generate a user's face from text description without their original photo. People and cards must not overlap; do not obscure subtitles. Subtitle text must match the voiceover; proofread proper nouns, numbers, and homophones. After removing pauses or changing speed, re-map the subtitle timing—do not directly copy original subtitle times. Keep subtitles within safe areas; check collisions at extreme moments of animation and zoom. For the ending, only use approved lines; a silent follow icon is acceptable; do not add narration or interaction promises without permission.
7. Audio-Visual Check and Final Delivery [Checkpoint 5]
Reference docs/quality-and-release.md. Target mix to -14 LUFS, true peak not above -1dBTP; measure and record actual values; do not claim compliance based on settings alone. A thumbnail overview does not replace watching the full video; watch and listen to the entire piece, check every transition and overshoot peak, and frame-by-frame check for overlapping materials or obscured subtitles. Automated checks only verify container/resolution/duration/loudness, etc.; do not claim they validate facts, audio quality, or visual occlusion. Deliver final-quality mp4, independently laid-out covers in 3:4/4:3/16:9, material sources, and editable project files; do not auto-publish or send messages on behalf of the user.
8. Publishing and Review
Prepare platform-optimized titles, descriptions, tags, and source attributions. Do not promise to automatically deliver resources after a comment unless the user actually has a method to do so. Publish at the time confirmed by the user; publishing is a separate authorization—do not upload simply because the video is ready. Approximately 48 hours after publishing, if the user provides real data, write three lines: where the drop occurred → evidence → what to change in the next video only. Prioritize metrics: click-through rate, 2-second skip, 5-second completion, average watch time, overall completion rate, saves, and new followers attributed to the work. Backend field names and definitions are in docs/publishing-and-review.md; leave blanks for missing metrics; do not compute when denominator is zero; do not draw simple causal conclusions across different lengths, topics, or accounts.
Public Package and License
Read ASSET_LICENSES.csv, THIRD_PARTY_NOTICES.md. Templates and scripts use MIT license; third-party assets do not automatically receive MIT license simply by being included in the project. The package only includes self-written geometric SVG placeholders; does not contain real person photos/stickers, cloned voices, original screen recordings, account data, private chat logs, or third-party fonts/music. Example final videos are silent flow demos, not real client cases, and do not guarantee traffic, production time, or commercial results.