A media workflow is a chain of custody.
Generation, editing, transcription, review, and release are different jobs. Treating them as one “AI video” button hides where rights, consent, files, and evidence can be lost.
This is an original workflow model, not a production report. No model, editor, renderer, microphone, or cloud service was installed or run. The examples are synthetic; no personal media, benchmark, cost, or quality result is claimed.
Five transformations, five review points
- 01
Brief + sources
Purpose, audience, authorized inputs, consent, exclusions.
- 02
Generate
Prompt, model identity, settings, references, raw output.
- 03
Edit + render
Composition, cuts, overlays, mix, export settings.
- 04
Transcribe
Audio scope, transcript, timing, correction.
- 05
Release
Approved master, credits, accessibility, provenance.
The durable artifact is not only the final video. It is the handoff record that answers which input became which output, under what permission, through which tool, and after which review.
Keep a rights ledger beside the timeline
| Asset | Permission question | Record | Gate |
|---|---|---|---|
| Synthetic shot | Were prompt inputs and references authorized? | Model/source, prompt digest, settings, output hash | Review identity, misleading detail, intended use |
| Narration | Did the speaker consent to recording, editing, transcription, and publication? | Consent scope, audio hash, edits, transcript revision | Authorized reviewer approves words and voice use |
| Music or stock | Does the license cover channel, term, territory, and modification? | Source, license snapshot, attribution, export linkage | Unresolved rights stop release |
Our editorial rule is to record separate, explicit permission for recording, editing, transcription, voice cloning, model training, and publication whenever a workflow uses those activities; one recorded permission is not treated as approval for all the others.
“Local” belongs to a stage
Local stage
Handy documents offline transcription; HyperFrames documents local preview and rendering. That does not establish where a coding agent, package registry, model download, optional provider, or later upload operates.
Service boundary
For every remote stage, record input bytes, destination, identity, purpose, retention assumption, and returned artifact. If unknown, label it unknown rather than private or secure.
Six release questions
- Does the final cut match the approved brief?
- Does every asset have recorded rights or an exclusion?
- Are identifiable voices and people within the agreed consent scope?
- Were captions, names, claims, and synthetic details checked by a person?
- Are corrected captions readable and timed?
- Can the master be traced to inputs, transformations, tool versions, and review?
What official sources establish
Generation
Wan2.1’s official repository documents generation tasks and distributed model/code artifacts. It does not establish a complete editorial or consent workflow.
Editable composition
HyperFrames’ official repository documents HTML/CSS/media compositions, linting, preview, and local rendering, while separating optional hosted services.
Roadmap is not runtime
OpenCut’s official repository calls the current project a rewrite and describes its API, MCP, headless, and scripting surfaces as coming work. They are not treated here as available production interfaces.
Transcription
Handy’s official repository documents offline speech-to-text dictation. It is not cited here as a timed-caption generator. Turning dictation text into captions still requires a separate timing, correction, and review step.
Sources checked. The workflow, ledger, boundary test, and release questions are original synthesis.
Download the blank media-rights ledger →