This project uses this local version and the Three.js production environment, and is not replaced by a global Skill of the same name. Preserve the author's conceptual ability, the user-specified style, and revision scope; a complete film is delivered on the basis of actual content, visuals, continuous action, and sound.
Each new film saves inputs, analysis, project files, and outputs in an independent working directory; sequels and local revisions follow the user-specified current project and qualified assets. When the user explicitly requests independent re-testing, recreate from the original video, SRT, explicit references, and current rules provided this time, without reading project files, finished films, analysis, matting caches, or historical conversations from other rounds, and without requesting old answers from other tasks; identical input is not automatically treated as a sequel. Public fonts, general mathematical tools, and authorized references may be used.
Applicable boundaries of the rules
The user's explicit requirements and valid revisions for this time take priority, followed by the specific requirements of the corresponding task in this project, and then general design methods are used to choose the implementation. Reference films, tutorial examples, and temporary exemptions from old films cannot cancel current requirements. "Default" means it must be executed when the user has not explicitly changed it, and the author cannot omit it on their own by stylistic judgment. Colors, fonts, shapes, and actions may be chosen according to content; whether to implement character outlines, protect must-read information, fully read stage rules, and complete delivery checks are not free options.
Live-action talking-head packaging must still prepare a synchronized independent character layer; matting and visible outlining are two independent decisions.When preserving the character's original video background, character outlining is prohibited; when replacing the original background behind the character, a clear, wide, and expansive stylized outline must be produced. Full-frame or small-window is judged by the same background condition, and one cannot automatically outline the person just because the window is shrunk. For specific background replacement, partial out-of-frame, and transition determinations, see uniformly Character Rules. After the character video is shrunk into a card, there must be a visible 3D floating layer, and the character within the frame must be re-centered for framing; container centering does not equal character centering. For the corresponding implementation, see Space and Camera and Container and Internal Framing. Local tasks such as fixing only the sound do not retroactively redo unrelated visuals.
This project's design organizes the image through space: clarify the front-back relationships among foreground, main focus, and background, and make the relationships visible through occlusion, scale, light and shadow, clarity, and motion; the finished film cannot be made into all elements pasted on one plane with equal clarity and equal weight. 2D materials and orthographic implementations may still be used, but the required spatial hierarchy must be implemented. Fonts for new designs must expand candidates and actually compare them in the complete composition, following Font Combinations and Posters , and cannot directly accept the default font.
Receive production input first
When SRT or live-action talking-head is involved, first use the local SRT Skill to complete content analysis. Before entering production, fully read its actual outputs, and verify the original SRT, original film/audio paths, and the user's latest requirements; do not start work based only on a chat summary from the previous round. Existing PLAN.md、timeline.json or qualified outputs such as preliminary content review are directly carried forward, and a new one is not recreated. Pure animation receives the user brief, real copy, and materials, and does not force the creation of an SRT.
Read fully before the stage, enter the next stage after completion
New complete films proceed according to the table below.Before starting hands-on work on each phase, read all pages of that phase in full as well as the technical modules already determined by the task; you cannot start just by reading the table of contents, the beginning, or a few paragraphs of each section. The corresponding detailed rules take effect from the first production run, not triggered by user corrections. Reading and production alternate; do not pour all subsequent phases into context first.
| Phase | Sections read in full | What is completed in this phase |
|---|---|---|
assets Materials and people |
Task boundaries, material semantics/authenticity, preservation of the original scene, synchronized cutout, person outline/stroke/container/framing, project foundation | Prepare real assets, synchronized person layers, and a usable project; decide the stroke toggle based on whether the person's original background is replaced, resolve source edges, and prepare the small-window secondary framing |
composition Scene and text layout |
Theme and space, opening, fonts and mixed Chinese-English typesetting, readability, safe area, original-scene backing light, before/after states, reading time, and sound conditions | Determine the visual direction and role of text for the whole film; compare fonts and key compositions within the complete frames of the opening and the most difficult relationship segment, reserve continuous action; do not lay out the complete film's static pages first |
animation Animation, camera, and transitions |
Action weight/duration/curves, group choreography, changes in person windows, advancement on the same canvas, camera, transitions, data evolution, and absolute time | First complete the continuous motion of the opening and the most difficult relationship segment, then complete composition, action, and adjacent transitions segment by segment according to semantics, expanding to the whole film |
sound Sound production |
Semantic anchors, sound coverage/intensity/timbre, action sounds, mixing, and protection of the original audio | Complete sound according to the original audio or music conditions; tasks explicitly specified as silent are handled according to their requirements |
review Full-film review and delivery |
Full text and callbacks, people, layout, camera and transitions, audio-visual synchronization, real output scope, and version entry points | First self-review and fix, then deliver the complete film and the actually verified scope |
These are the production phases of a single author, not dividing the film among different authors. The completion condition for composition is that the visual direction, text system, and representative key compositions hold; it does not require every other segment to be laid out as a final draft first. Then, only after reading all pages of animation in full and the applicable modules, does formal animation choreography begin; you cannot read the first page and then write the whole film's animation, then go back and read the remaining pages. The opening and the most difficult relationship may belong to the same continuous segment; choose according to the actual content. The sample segment must include the process of reveal, reading, and connecting to adjacent content; only after fixing it do you expand, and you cannot expand based only on a still frame or a successful export. Each subsequent semantic segment completes composition, motion, and transitions together, following the typesetting rules already read in full; it is forbidden to apply a uniform full-page slide-in/fade-in function to the whole film. The semantic timing of the original audio constrains the visuals from the input onward, and music-led tasks also obey the real rhythm from composition onward; the sound phase is responsible for actual production and mixing. This process does not add user approvals or phase records.
Use the paginated entry point to read the original text in full
Execute from the project root:
node scripts/read-stage.mjs --stage assets --kind talking-head
For pure animation use --kind animation. The script follows the phase source list to return the complete body text/complete chapters of the original file, with no single page exceeding 2200 body characters, preserving source and page numbers; it is not a rules summary. The list checks all reference chapters of both Skills, explicitly errors when newly added/renamed chapters are unassigned, and does not silently skip reading.
- Each tool call outputs only one page; execute the returned
NEXT_READ, until you seeEND_READINGDo not loop-print all pages in a single tool call, and do not mix in source code, directories, image data, or probing JSON. It is recommended that a single tool output budget be at least 4000 tokens, and give the outer tool the same amount of space; the character limit is not an exact token estimate. END_PAGEorEND_READINGis missing, or the tool reports truncation/failure, reread that page and fill it in; production cannot begin. The next page carries the source hash; if the material changes midway, reread from the current stage to avoid mixing old and new versions.- After the page arrives complete, carry out the work of this stage. The last page of the script gives the next-stage command after completing the work; continue executing it, without waiting for the user to remind or approve stage by stage; when the user explicitly requests reviewing a certain stage first, comply.
- Before starting work in each production stage, compare against
--listtask/implementation trigger conditions, and find all modules applicable to this stage. When a task has already determined that a certain implementation is to be used, before starting work use--module 名称to read the corresponding module in full; module names and trigger conditions are in--list. For example, before using the Three.js composition adapter readthree-framing; for interactions such as drag, crop reveal, and selection readui-interaction, and for producing solid UI surfaces additionally readui-surface. These are implementation dependencies, not remedial entry points after an error occurs. - When the SRT output or the selected template analysis is long, you may use
--file 文件路径to read it fully in pages. The reference index is for lookup; selected cases must be read in full for analysis and the corresponding continuous actions actually viewed; search results must not be treated as already-read cases.
Cross-file links within references indicate the source and are implemented by the current stage checklist; do not follow the overview of an old entry point to recursively preread later stages. The wording "trigger/fix" in existing references is used to explain applicable situations; rules classified under the current stage must still be read in full. User-authorized local modifications enter from the affected stage, fully read the corresponding content of that stage, and check adjacent impacts; for sound-only changes, reuse the already-confirmed visuals. Do not use local revisions as grounds to expand into redoing the entire film.
Context and execution boundaries
Between stages, directly continue using the SRT output, current source code, assets, and previews; do not add stage check-ins, design decision tables, or inspection ledgers. Sources already read in full by a stage checklist do not need to be reread in full just to follow links; content required by the next stage checklist must still be read in full.
Pagination solves output truncation, and stage arrangement reduces early reading and repeated input;it does not delete already-occupied context, nor does it guarantee that a long film can be completed in a single window. If platform context compression occurs, recover the current work position from existing analysis outputs, the project, and the most recent preview; when it cannot be confirmed that the current stage was read in full, reread that stage, and do not treat the compression summary as the original rule text. There is no need to reread all historical conversation or restart completed production.
The script only provides a complete, locatable reading order; it is not a background production scheduler; completing the reading does not mean the actions or visuals are qualified. Check the actual delivered version according to the review stage, and fix already-discovered problems such as erroneous strokes on original scenes, missing strokes after background replacement, cards stuck to the bottom, improper character cropping, text occlusion, and repeated page-flip-style animation. When the necessary viewing/listening has not been completed, you may only mark the preview as pending inspection and specify the exact gaps; you must not use a single limiting statement to declare it a finished film; when the user explicitly requests an uninspected raw draft, deliver the raw draft within that scope.
For version sources see installation records; the original production principles are incorporated by chapter into production principles, not another entry point that requires full prereading.