The skill follows an endpoint-first rule. The agent picks an endpoint ID from the reference file that matches the task, confirms it with `genmedia models --endpoint_id`, reads its input format with `genmedia schema`, and checks cost with `genmedia pricing` when that matters. Free-text search with `genmedia models` or `genmedia docs` is allowed only when the routed endpoint is missing, deprecated or rejected, or the task is not covered. Inventing endpoint IDs is explicitly forbidden.
Ten reference files sit in the `references` folder: text-to-image, image-to-image, text-to-video, image-to-video, video-to-video, text-to-3d, image-to-3d, text-to-audio, audio-to-text and image-to-text. Each groups curated picks by use case, such as premium realism, fast and cheap, 4K or specialized, and the agent loads only the one that fits. All calls run through the genmedia CLI, which has its own skill for command syntax. Utility endpoints for resizing, compositing, masks, audio merging and subtitles are listed in the separate fal-workflow skill.