Deliver an actual video that viewers can follow to derive and operate. Continue the style, language, and parameters the user has confirmed; the knowledge topic, number of chapters, coordinate conventions, thresholds, and data all come from this task, and do not carry over the business settings of an old video.
Standalone use and optional enhancements
This skill completes video production independently and does not depend on other skills. If the skill list provided by the current environment includes awesome-docs, complex content organization may reference it as needed; if it includes shuorenhua, the spoken script may be polished as needed. Read through the actual paths provided by the list; do not assume adjacent directories or fixed installation locations. If they are missing or the required reference files do not exist, skip the enhancement and continue with this skill; do not scan other agents' installation directories, and do not trigger installation, updates, or installation prompts. If a skill the user explicitly names fails to load, briefly explain and do not claim to have executed it.
Video explanation and visual rules
- Storyboard in the order viewers understand: first establish the problem, inputs, and necessary concepts, then demonstrate changes, calculations, and results. Each scene focuses on one question and retains intermediate derivations; do not move document paragraphs page by page into the video.
- Division of labor between visuals and narration: narration explains what is happening to the current object and why; the visuals show key points, lines, surfaces, and numerical changes. Avoid the narration already moving to the next step while the visuals are still on the previous step. The visuals retain short titles, necessary labels, formulas, and subtitles, and remove repeated explanations.
- Use LaTeX for formulas: formulas inside and outside the figures, fractions, vectors, matrices, and sub/superscripts are typeset via a math renderer; check again after dynamic replacement and final export. Chinese explanations are typeset separately, and variables, units, symbols, and geometric calculations remain consistent.
- Use Arabic numerals for subtitle values: quantities, measured values, versions, ranges, and percentages use Arabic numerals; natural words and verbatim quotations are retained. Display text and spoken speech are stored separately; do not perform context-free Chinese numeral replacement. Subtitles go in an independent safe area and are burned into the video by default to avoid duplicate display with external subtitles.
- Diagrams serve the explanation: use 3D for spatial relationships, and 2D for projections, curves, and processes as needed. First explain the origin, reference frame, and objects, then highlight key geometry along with the explanation; other necessary objects remain dimmed and retained. Zooming in or rotating the view does not change measured values; the camera viewpoint and the physical camera within the scene are labeled separately.
- Spoken narration is direct and accurate: clearly state the object, action, conditions, and result, and remove vague previews and repeated summaries; retain facts, terminology, numbers, and uncertainty, and do not mechanically rewrite sentences according to a banned-word list. When reading through, check whether premises precede conclusions and whether viewers can understand the narration while following the current visuals.
Workflow
1. Define content and evidence
First find the existing project. When modifying a final video, keep the old version; when only changing subtitles, do not also change algorithms or values. Distinguish user-specified values, source values, derived values, and instructional constructed values; clearly label unknown parameters. Material production does not modify production code, configurations, or device states.
Record for each act: the questions the audience needs to understand, prerequisite concepts, inputs/outputs, data used, graphical actions, formulas, displayed subtitles, and voiceover. Check the entire derivation and causal chain first, then write the animation. When compensation or control is involved, distinguish prediction, actual execution, re-measurement, and criteria; organize other knowledge according to its own logic.
For geometry- or computation-intensive tasks, read as needed references/geometry-and-reasoning.md。
2. Copy, voiceover, and timeline
Read through and polish the entire copy first, then generate speech. Explicitly store the displayed text text and the spoken text speech for each segment; number expansion changes only the reading, not the value, unit, sign, or version meaning.
Use real word boundaries or documented forced/manual alignment to generate subtitle timings. Subtitles and key animations share the timeline; after changing narration, regenerate invalidated audio and boundaries, and update event anchors. When only changing an equivalent-meaning display format, unchanged voiceover caches may be reused, but subtitles must still be rebuilt.
For new projects, subtitle data structures, speech services, export, and validation, read as needed references/production.md。
3. Implement and deliver
Prefer reusing an existing project; switching frameworks is not mandatory. When there is no project, you can use the included Motion Canvas / Three.js starting point:
python3 SKILL_DIR/scripts/scaffold.py /absolute/path/to/video-project
This scaffold only copies files and generates a silent layout sample; formal voiceover, dependency installation, and rendering are performed separately. The default is 1080p / 30 fps, adjustable as needed for the task.
Before delivery, check computational invariants, render logs, the final MP4 audio/video and subtitles; spot-check key actions and formulas in the final decoded frames, not just the web page or static storyboards. Record manual audiovisual spot checks separately from automated pixel/audio checks, and do not overstate coverage.
Return the actual MP4, with a chapter player, subtitles, and editable source project attached as needed. In Codex, display using absolute-path media embedding. Exclude runtime, temporary large files, and sensitive configuration from packaging; do not publish, commit, push, or install globally by default.
Purpose of the attachments
SKILL.md is the only required file. This skill additionally retains 2 on-demand reference documents, 1 copy script, 1 piece of interface metadata, and 19 runnable template files. The dependency locks, render entry points, audio alignment, and validation code in the templates are for reproduction; when using an existing project, there is no need to read or copy them. Specific business videos, measurement data, finished videos, caches, and production logs stay in the project and are not placed in the general skill.