specialization:media
Media Processing reference
Generative media point tasks — one per modality and operation — plus one end-to-end
media
Generative media point tasks — one per modality and operation — plus one end-to-end production pipeline that carries a brief through to published assets and post-publish metrics.
Point tasks
— Image-generation persona. Parse creative brief -> select optimal model (Imagen 3/4, Flux, DALL-E, Stable Diffusion) -> generate variants in parallel -> validate technical + creative quality -> organise outputs with metadata.
— Image-editing persona. Analyse source + operation request -> select tool (Imagen Edit, DALL-E Edit, Stability Edit, Photoshop AI, Upscaler) -> apply operation (inpaint | outpaint | object-removal | background-replace | style-transfer | upscale) -> validate edge quality / color consistency / artifact absence.
— Video-generation persona. Parse request (text-to-video | image-to-video | video-to-video) -> select optimal model (Veo 2/3, Luma, RunwayML, Stable Video, Minimax) -> generate via MCP GenMedia with camera/lighting/composition parameters -> validate technical + content quality -> retry with fallback model on low-quality outputs.
— Video-editing persona. Analyse source + request -> select tool (Veo Edit, FFmpeg AI, DaVinci Resolve, RunwayML Edit, Video Enhance) -> run per-op pipeline (temporal-inpaint | stabilise | color-grade | upscale | transitions | scene-cut | audio-sync) -> validate frame consistency + audio sync.
— Speech-generation persona. Analyse text + voice requirements (language, style, emotion, SSML) -> select model (Chirp 3, Azure Speech, ElevenLabs, OpenAI TTS, AWS Polly) -> synthesise via MCP GenMedia -> validate naturalness / pronunciation / audio specs.
— Music-generation persona. Parse composition brief (genre/mood/duration/instruments) -> select optimal model (Lyria, MusicLM, AIVA, Mubert, Amper) -> generate via MCP GenMedia -> apply mastering + stem separation if requested -> validate musical coherence + technical audio.
image-generation.js(@process specializations/media/image-generation)image-editing.js(@process specializations/media/image-editing)video-generation.js(@process specializations/media/video-generation)video-editing.js(@process specializations/media/video-editing)speech-generation.js(@process specializations/media/speech-generation)music-generation.js(@process specializations/media/music-generation)
Pipeline
media-production-pipeline.js (@process specializations/media/media-production-pipeline) — End-to-end media production: brief -> research -> script -> produce -> review gates (editorial, legal, brand) -> publish -> measure.
Its task ids, in execution order:
| Task id | Kind | Role |
|---|---|---|
media.research | agent | Research the brief |
media.script | agent | Script/outline from the research |
media.produce | agent | Produce the asset |
media.review-gate | breakpoint | One pass per gate (editorial, legal, brand), up to 3 attempts each |
media.publish | agent | Publish to each entry in publishTargets (in parallel) |
media.measure | agent | Collect 24h post-publish metrics |
Known gap (OPEN)
The pipeline's review gates are plain ctx.breakpoint calls, **not** routedBreakpoint policy gates: there is no policy-gated tag, no actionId, and no gatedActions audit record of the decision. Worse, each gate is skipped outright when the matching reviewers[gate] input is absent, and media.publish carries no approval gate of its own — so a run with no reviewers publishes with zero human review. This is stated here as an **open gap, not a fixed behaviour**; adding routed policy gates is a separate change and is not done in this pass.
Assets
— the brief-to-prompt convention shared by the six point tasks.
- Skill: `skills/generative-media-prompting/SKILL.md`
- There are no agents in this specialization.
---
Descriptions in this README are transcribed from the files' own @description headers, not invented.