What problem does it solve? Calling generative media models fails silently in ways text-model habits do not predict: seeds that do not reproduce, masks that do not protect pixels, raw PCM saved as unplayable .wav files, transcripts that invent sentences over silence, and provider URLs that expire after the job completes. This Skill encodes the invariants, job lifecycle rules, and provenance obligations that make generation pipelines correct, affordable, and accountable. ## Core Features & Use Cases - Generation and editing rules: Resolve models at runtime, read per-model schemas, decide generate-versus-edit by what must survive, treat masks as prompt-level hints, and composite deterministically when pixel-exact regions are required. - Job lifecycle and cost control: Choose synchronous, polling, or webhook transports, persist job ids before polling, avoid non-idempotent retries, verify webhook signatures, and compute batch cost before submission. - Speech and transcription handling: Chunk TTS input on sentence boundaries, keep voice configuration identical across segments, wrap raw PCM in proper containers, split ASR input on silence, and re-add chunk offsets to timestamps. - Reproducibility and provenance: Manage generator state, CPU versus GPU noise, version pinning, C2PA manifests, SynthID watermarks, disclosure obligations, and generation records linking assets to the calls that produced them. - Use Case: A team re-running approved image prompts on a different GPU gets different results despite identical seeds; the Skill identifies the reused generator state, the CPU/GPU RNG difference, and the missing version pins, then proposes archiving approved bytes as the artefact of record. ## Quick Start Ask the agent to review your image generation or narration job using the generative-media skill and report lifecycle, cost, and provenance findings.
