game-creator: images, video, and audio in an agent workflow

The work continues after generation

A game prototype may need character images, video, narration, and background music. Calling a model is only part of the work: requests can become uncertain, downloads can fail, and long jobs may need to resume.

game-creator organizes these steps in a Skill and command-line tool. Version 2.0 covers image generation and editing, image understanding, text-to-video, image-to-video, speech, and background music.

One entry point, several protocols

Media APIs do not share a single request format. Adapters distinguish gateway JSON from multipart image edits, and separate video protocols. Unknown video models require an explicit adapter.

Users configure their provider and model, or override the model with --model. Whether a particular model works still depends on the capabilities available through that provider.

Inspect the plan first

The tool requires Node.js 22+, has no third-party runtime dependencies, and supports Windows, macOS, and Linux. After following the installation guide, run these in the Skill directory:

node asset-gen.mjs --help
node asset-gen.mjs doctor
node asset-gen.mjs image "pixel art fantasy character" "outputs/character.png" --dry-run

Dry runs validate and display a plan without making a generation request. The doctor command checks configuration; it does not verify account balance or live model availability.

Resume a long task

Video generation may involve submission, polling, and downloading. With --submit-only, the tool saves task state. A later resume command continues polling or downloading instead of submitting another generation.

node asset-gen.mjs resume "outputs/clip.mp4.task.json" --json

Generation POST requests are not automatically retried. An uncertain network response could already represent a paid task. Keeping its state helps separate recovery from a second purchase of the same generation.

Check the output before delivery

The tool rejects HTML error pages, invalid media responses, and files whose type does not match the extension. Media downloads do not receive the generation API credential, and existing files are not overwritten.

Container validation does not establish creative quality. Dimensions, duration, character consistency, images, and audio still need inspection. Explicit fallback choices also keep local background music distinct from remote model output.

Current limits

Tests use local mock services to check implementation behavior. They do not establish that every paid provider works. Mask editing, automatic conversion, overwrite operations, and concurrent writes to one output path are not supported.

The source is public on GitHub, but the repository currently has no open-source license. Check its current terms before reuse or redistribution.