Text-to-Video Workflow

What does Text-to-Video Workflow do?

Use the project directory selected by the user as the production root directory, and comply with applicable project instructions and long-term user feedback. First, inventory existing renderers, fonts, voice, and audio tools. The repository provides initial animation components and a working full chapter example; reuse these first, then add scenes needed for the content.

The user provides the original text, link, or topic in one go; the Agent is responsible for all creation, checks, and fixes below, and does not push back approval for each shot or sample reel to the user. Only raise specific gaps when body text, assets, or genuine authorizations are needed. Do not generate fake successful output when no voice service or factual evidence is available.

Scope & Specifications

• Default task-short: 9:16, 1080×1920, 30 fps, 60–90 seconds, 5–8 shots, H.264, video bitrate at least 8 Mbps, AAC voiceover, real voice-aligned SRT, material source and cost log. Task acceptance also requires editable projects for CapCut/Premiere Pro/FCP and an open instruction; Remotion source code does not replace this.
• Use extended when the user explicitly requests something longer or follows a three-minute format: 165–195 seconds. Do not pass off an extended version as a task-shot acceptance.
• Use chapter-proof when the user first validates a long-form chapter: 90–150 seconds, 1920×1080, 30fps, H.264/YUV420p, AAC, SRT. The current example uses CRF18; do not replace visual inspection with a fixed high bitrate. Read the long-form chapter workflow before doing chapters or long-form pieces. Full 8–10 minute quality has not been validated; do not automatically extend the current duration.
• Fixed visual baselines and review methods are described in production specifications. Primary color, font, subtitle area, sound, and motion grammar are fixed; content and explanatory actions vary with the article. The default style baseline is in style-lock.json; adjust and lock according to the user's existing brand requirements; the configuration itself does not prove aesthetic approval.
• Always no external distribution. Preview with no-audio MP4, paused by default; save the voiceover version separately. Deliver to the user-specified directory; public release or distribution to others requires user authorization.

One Entry Point, Internal Iteration

1. Input & Evidence: Save the original text and its source scope. If a link is unreadable, do not fabricate the author's viewpoint based on the title; an independent research question may be set, but clearly mark source boundaries. Record key arguments, figures, qualifiers, original calculations, and first-party evidence.
2. Content Design: Form a specific question, a mechanism, a set of testable examples, and applicable boundaries. Short film defaults to a single mechanism; extended version allows a more complete argument. Synthesize voiceover only after fact-checking.
3. Story & Action: Select fixed grammar actions such as comparison, regrouping, cumulative, filtering, etc., based on the content. Write each shot's object identity, start and end state, the real time when the action occurs, and how the next shot picks up. Do not just change titles, icons, or bottom captions.
4. Reuse Production: Reuse TTS, WordBoundary, fonts, Remotion, and FFmpeg from recent successful projects; write only the scenes needed for the content. Do not rebuild the base engine, and do not copy node_modules junctions. Synthesize the voiceover as a whole segment; realign the timeline after changing the script. Cache key includes manuscript, voice, and speed.
5. Internal Validation Loop: First render 8–13 seconds including explanatory actions and transitions. Check at mobile size, consecutive frames, text, object state, numbers, and lip sync. Explicitly list timecode issues, fix locally, then review again. Then render the full film and check the final encoded file. At most three rounds of similar local fixes; if still unqualified, explicitly mark as pending fix, do not fake qualification.
6. Delivery & Status: Run the acceptance script and retain itemized status. Separate technical checks, Agent visual review, and user aesthetic approval. Deliver voiceover MP4, silent preview, SRT, editable source code, segmented frames, separate audio, sources, and logs. When the user has explicitly verified the editor manually and requires no re-verification, record the user confirmation and scope, then proceed directly with production; do not start CapCut/PR again, and do not fake that the Agent opened a new project. Mark XML as exchange timeline.

Music & Assets

Select assets by purpose: reference materials provide evidence, calculations explain mechanisms, illustrations establish scenes; record source, license, usage interval, and related argument for each. Public examples include fonts, voiceover, clean segmented frames, limited paper heading excerpts; do not include full text, complete papers, original reference videos, or unauthorized redistributed music in the public package.

When the user specifies a reference music track, first confirm the full video duration, music track duration, and real audio content. A 6-second file marked as "original sound" by the platform only proves a 6-second segment was obtained, not that the original track is only 6 seconds. If the source track is insufficient, continue searching; do not use a voiceover original track as pure BGM. Loop only when the user accepts segment reuse; do not automatically replace full music with a short loop. Once usable music is obtained, read the sound and materials workflow and use the local mixing tool.

Reusable Entry & Real Capabilities

Use the user's existing or explicitly chosen production project. Remotion, Canvas, or other deterministic rendering tools may be used; specific voiceover and export commands are provided by that project. The directory and JSON conventions for the auditor are described in the input protocol. source/script.json is the Agent-adapted input, not a model that automatically turns arbitrary text into shots.

This Skill is an Agent orchestration entry point, not an unattended deterministic full-text-to-video program. Topic selection, storyboards, and visual review are still executed by the Agent; the user does not need to command step by step. Only when three new articles not involved in debugging run through successfully and the fix log is complete can the pipeline for a fixed scope be reported as verified; do not promise first-pass perfection for arbitrary articles.