Theme-First Subtitle Motion Design · 2026-09-27
Treat subtitles as short-duration motion pieces that carry real text. The user's theme and current requirements take priority; subtitles only constrain text accuracy, actual screen readability, time and position control—they do not mandate long rectangles, thick fonts, heavy frames, or abstract UI. First conceive a thematic event and recognizable composition, then adapt short or long sentences. Simplicity, abstraction, concreteness, characters, and installation are all valid approaches. Do not elevate a single failure or old preference into a global ban.
Theme names may use related symbols, patterns, objects, characters, and spatial relationships; no separate authorization for visual design is required. Select a few highly recognizable elements and compose them with text/support structures. Avoid random small icons, color swaps, or titles claiming to resemble a game. User instructions like "no characters" / "no frames" / "retain character" apply only to the current task. When no such requirement is given, neither force characters nor preemptively forbid them. The presence, thickness, non-rectangular outline, open boundaries, and proportions of the design base are determined by readability and the theme.
The planning description must briefly include: visible theme basis, relationship of body/support/logo, entrance event, activity during reading, and exit event. Theme identification check: Without the name or description, can the direction still be identified from silhouette, pattern, or motion? After converting colors to grayscale, does a unique form remain? Ordinary slabs / green-gold borders / bold white text are only baseline designs; do not replace the design with a name like "Temple/Ancient".
Actions are arranged causally: trigger moves first → structure responds → text forms → local activity → dissolve/transform/recycle. Each work selects appropriate events; do not uniformly apply rotate, sweep, pixelate, or fade-in. Opacity may assist visibility; full-group opacity or slight translation alone does not constitute a thematic action. Both entrance and exit involve visible changes to the main body; the exit may differ from a reverse of the entrance. During the reading phase, the body text should be clear; peripheral meaningful activity is allowed. Short cues reduce secondary events and travel, retain theme identification, and do not pack the entire scene by accelerating everything.
Text design follows the theme: choose real Chinese fonts, consider stroke temperament, hierarchy, texture, and letter spacing; do not force stroke/bevel on every draft. Without a key word, the work is still complete. Short sentences adapt by tightening naturally; long sentences wrap naturally. Do not alter the body text or use extra English filler slogans.
Use the host's declared HTML/SVG/CSS static layers and unified MotionSpec clock; security, text binding, font loading, reproducible seek, and export consistency are hard technical boundaries. Do not interpret these technical boundaries as aesthetic limits. Free source is the normal creative path; components are only optional quick materials; do not treat unsuitable presets as a fallback for failure.
New works are freely conceived; revisions of existing works first clarify what is locked for this task. Preserve structures and actions the user has approved. Failed drafts the user has rejected serve only as negative examples; they are neither automatically frozen nor automatically overwrite accepted works. Technical acceptance is through preview/save/adopt; aesthetic suggestions do not trigger unauthorized paid rework. Actual checks cover normal proportions, details, full normal-speed animation, 6 and 1.5 seconds, sentence changes, forward/backward seek, and clear begin/end; technical pass, author review, and user satisfaction are recorded separately.
Project Collaboration & Effective Path
New works are freely created according to current requirements; enhancements only protect what the user has actually approved/explicitly locked. Failed drafts are not a satisfactory baseline. Strict material revisions do not alter structure or action; explicit refactoring allows collaborative changes; do not confuse the two types of tasks. Keep old drafts, model first drafts, and local refinements saved and identified separately.
Actual generation loads assets/design-skills/abstract-caption.md, material-caption.md, and any matched theme guidance; design_guidance returns the version and SHA of each file. The project skill stays in sync with the in-package copy; after modification, check the actual prompts for normal planning, implementation, and revision—do not only edit this file. Security/font/binding/clock boundaries are governed by code capability contracts; do not redundantly stack old restrictions.