Knowledge Fable Video
Produce the episode, not merely a plan. A complete manuscript is sufficient input. Infer ordinary production choices from the text and available tools; ask only when a missing choice changes the concept, factual claim, language, or delivery target.
Preserve and structure the source
- Keep the supplied manuscript unchanged as the source artifact. Put editorial structure, narration, and visual adaptation in separate files.
- Identify the hook, setup, pressure, irreversible choice, delayed payoff, concept reveal, story-to-concept mapping, and boundary.
- Verify unfamiliar or time-sensitive claims with primary sources before adding them to narration.
- Tell the story before naming the formal concept. End with a clear definition, the mapping, and at least one limitation or countercondition.
- Prefer a final runtime below 8 minutes and treat
8:00as the normal planning ceiling. Preserve the source manuscript, but compress long production narration aggressively: keep the central question, causal chain, essential turning points, payoff, formal definition, story mapping, and boundary; merge or remove repeated arguments, parallel examples, secondary branches, ornamental detail, and long transitions. Do not substitute faster narration for editing. Exceed 8 minutes only when further compression would distort the core claim; use the shortest responsible version and state the deviation.
Build narration and the timeline
- Split narration into short spoken cues at natural changes of action, subject, or argument.
- Generate or record one audio file per cue. Use measured audio duration as the authority for caption and scene timing.
- Before full synthesis, review names, places, technical terms, abbreviations, symbols, dates, numbers, and genuinely ambiguous polyphonic characters.
- For Chinese narration, never create a pronunciation override containing
地or得. Leave both characters to the TTS engine's contextual reading; forced phrase overrides containing them can distort prosody. - Keep any other pronunciation fix at the smallest useful phrase level, verify the effective reading, and listen to every affected cue before synthesizing the complete voice track. Do not put pinyin into visible captions.
- Keep narration dominant in the mix and change music only at meaningful chapter boundaries.
Design and generate visuals
- Write a scene plan before generating assets. Record each scene's purpose, key subject, shot count, palette, motion, and caption-safe negative space.
- Establish stable character anchors and one reusable visual baseline. Generated images must contain no captions, labels, titles, or watermarks.
- Prefer a coherent editorial storybook language across both story and explanation. Static illustrated plates plus restrained compositor movement are the default.
- Reuse a plate when only attention, timing, camera drift, or light changes. Add another shot only when the spoken beat changes subject or location.
- Keep symbolic colors, objects, and outcomes hidden until the story introduces them.
- Render all exact text locally in the compositor.
Integrate and finish
- Isolate every episode's source, narration, audio, visuals, manifests, code, and outputs under an episode-specific slug.
- Use one compositor as the owner of the final timeline, subtitles, sound mix, transitions, and render.
- Keep concept scenes in the story's visual world. Reveal the question, formal name, mapping, mechanism, and boundary in stages; narration carries the full explanation.
- Build a matching cover from a no-text background plus locally rendered typography. Check it at full size and thumbnail size.
- Deliver the final encoded video, cover, one recommended title, a short introduction, and focused tags in the requested language or the manuscript's language.
- Verify the actual encoded video through complete decode, representative cross-scene frames, audio continuity, captions, and flagged pronunciations. Compilation alone is not completion.
- Stop before upload or publication unless the user explicitly authorizes that external action.