a5c-ai Generative Media Prompting
All six media point tasks open with the same four-step shape, visible verbatim in their
@description headers. Writing it down once removes six copies of the same tacit
convention.
The four steps
- Parse the brief. Extract what the request actually asks for — the creative intent for a generation task, or the source asset plus the requested operation for an editing task.
- Select the model or tool. Choose the model (generation) or tool (editing) that fits the parsed brief. Each point task names the candidates it selects among; the selection is part of the task, not a caller input.
- Construct the prompt. Turn the parsed brief into the prompt (and, where the task supports them, the structured parameters that accompany it).
- QA the output against the brief. Validate the result before returning it. Every point task ends in a validation step, and what it validates is modality-specific.
Per-modality notes
Limited to what the existing files already state:
image-generation.js— generates variants in parallel; validates technical and creative quality; organises outputs with metadata.image-editing.js— selects among named editing tools; validates edge quality, color consistency, and artifact absence.video-generation.js— the parsed brief includes the request mode (text-to-video, image-to-video, video-to-video); prompt construction carries camera, lighting, and composition parameters; a low-quality output is retried with a fallback model.video-editing.js— selects among named editing tools and runs a per-operation pipeline; validates frame consistency and audio sync.speech-generation.js— the brief includes language, style, emotion, and SSML; validates naturalness, pronunciation, and audio specs.music-generation.js— the brief includes genre, mood, duration, and instruments; mastering and stem separation are applied only if requested; validates musical coherence and technical audio.
Scope
This skill describes prompt construction only. Publication, review, and licensing
decisions are out of scope and belong to
../../media-production-pipeline.js. No model lists
or vendor guidance beyond what the point-task files themselves name.