I Built an Editable Timeline for AI Agents: Skills, MCP, and Transactional Video Editing

Описание и применение

Создана система редактирования видео, где агенты могут последовательно изменять проект, сохраняя все метаданные и обеспечивая проверяемые планы изменений.

Задача

AI agents can generate FFmpeg commands, but they lack a way to edit a project incrementally, preserve state, and provide a reusable, inspectable plan for subsequent edits.

Результат

The architecture enables agents to perform incremental, reversible edits on a video project while preserving all necessary metadata, enforcing preconditions, and providing a dry‑run diff before applying changes. It supports idempotent retries, revision checks, and transactional application to avoid partial edits.

Рабочий процесс

• Inspect project state
• Agent analyzes media and generates a structured edit plan
• Validate plan against project preconditions
• Generate a semantic diff to preview changes
• Apply the plan transactionally to produce a new project archive
• Re‑inspect the updated project

Доказательства

A rendered MP4 does not preserve the decisions needed for the next edit: clip identity, source ranges, timing, transforms, and caption/audio relationships.

Источник

Timeline Studio uses an editable .timeline project as the source of truth.

Источник

The workflow is: inspect -> plan -> validate -> diff -> apply -> inspect

Источник

The local CLI and MCP writes also require a new output path and reject overwriting the input or an existing output.

Источник