Flow Video Producer Skill

An AI video production skill for agents with browser automation, including Claude Code,
OpenCode, and Codex environments. The agent plans the video, operates Google Flow to
generate images and video segments, then assembles downloaded media locally.

How it works

1. Define visual direction in philosophy.md and style.json.
2. Plan scenes and segments in scene-breakdown.md.
3. Write prompts, output paths, and status into pipeline.json.
4. Generate reference assets.
5. Generate a starting keyframe for each scene.
6. Generate segments sequentially, using the previous segment's last frame as the next starting image.
7. Assemble scenes and deliver output.mp4.

A scene represents a narrative/camera unit. A segment represents a generation-sized
chunk (8 seconds by default; verify current Flow options). A 20-second continuous scene
can contain 8 + 8 + 4 retained seconds. Frame chaining helps visual continuity; review
joins for motion and character consistency.

The main agent handles planning, review, status updates, and assembly. Generation role
instructions ship inside the skill and are passed to generation sub-agents through
the host's delegation tools. Tasks sharing one browser session run serially.

Setup

Requirements: browser automation tools, a signed-in Google account with Flow access,
and Python 3.9+ for local media processing. Generation uses Flow's cloud service; it
requires neither a local GPU nor a Google API key. Verify generation settings in the
current UI. Default models are Nano Banana Pro for images and Veo 3.1 Quality for
video, with one output per prompt.

For Claude Code:

/plugin marketplace add zysilm-ai/flow-video-producer-skill
/plugin install gemini-video-producer@gemini-video-producer

If MCP Playwright is missing in Claude Code:

claude mcp add playwright -- npx @playwright/mcp@latest

Restart the host if needed to load its browser tools. In other hosts, configure their
available browser tooling instead of running Claude-specific setup commands.

For manual installation, copy the entire skills/gemini-video-producer/ folder into
the host's skills directory. Its scripts, references, and example style are self-contained.

Install local dependencies in a virtual environment:

python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt

Then describe the video, for example: “Create a 24-second first-person battlefield
experience.” The skill establishes creative direction and plans scenes before generation.

Local assembly

# Assemble completed segments from the project manifest.
.venv/bin/python scripts/merge_videos.py --pipeline output/example/pipeline.json

# Manual assembly, including a single source if needed.
.venv/bin/python scripts/merge_videos.py -o scene.mp4 seg-A.mp4 seg-B.mp4 --durations 8 4
.venv/bin/python scripts/merge_videos.py -o output.mp4 scene-01.mp4 scene-02.mp4 --transitions dissolve

# Extract the last frame of the retained interval for chaining.
.venv/bin/python skills/gemini-video-producer/scripts/extract_frame.py seg-A.mp4 -o after-A.png --duration 8

Audio is preserved by default; --no-audio requests silent output. cut and fade
preserve the sum of scene durations. dissolve and wipe overlap scenes and shorten
the final video by the overlap duration. Fade goes through black; wipe reveals left
to right. The manifest records transition duration, and the skill explains runtime
planning in pipeline.md.

The helper emits JSON on success or failure and returns a corresponding exit code.
Pipeline assembly marks completion only after successful output, and keeps originals.

Repository structure

.claude-plugin/                  Claude plugin distribution metadata
skills/gemini-video-producer/
  SKILL.md                      Portable production instructions
  references/                   Generation roles, schema, and production guidance
  scripts/                      Assembly, extraction, and dependencies
  assets/example-style.json     Visual style example
scripts/merge_videos.py          Repository compatibility entry point
requirements.txt                Repository local processing dependencies
tests/                          Synthetic media integration tests

Run local verification with .venv/bin/python -m unittest discover -s tests -v.
Browser generation requires real Flow access and credits; local tests use synthetic
media and do not exercise that external service.

MIT license; see LICENSE.