media-gen

name	media-gen
description	Generate images, video, audio, or other media via external AI providers, with Claude authoring the prompts and orchestrating the calls. Use when the user asks to generate/create an image, video, audio, icon, or asset, refine a generation prompt, or wire a media/AI provider. Provider-agnostic: works with any provider that has an adapter + an API key in .env (Google, OpenAI, and others — not a fixed list). Claude writes the prompt, calls the API, saves the asset to assets/, and logs cost. Never prints API keys.
license	MIT
allowed-tools	Read, Write, Edit, Bash
argument-hint	[image|video|audio] <description>

Media generation — provider-agnostic, key-driven

Claude is the prompt author and orchestrator; the media comes from an external provider's API (Anthropic has no image/video generation — when "Anthropic" is chosen it means Claude writes the prompt and another provider renders). This is one consumer of the genesis integrations registry: a provider is just an adapter

• an API key.

How it works

1. Pick the provider from those enabled (an adapter in integrations/registry/ with category ai-media and its *_API_KEY present in .env). If several fit the requested capability (image/video/audio), pick the best per the seed's default / the registry, log the choice in docs/DECISIONS.md, and proceed — don't ask (A5). Never assume a provider that has no key configured.
2. Author the prompt. Turn the user's request into a strong, specific generation prompt (subject, style, composition, constraints, negative prompts, aspect ratio, duration for video). Record it in the asset's sidecar so it's reproducible; proceed — don't stop to confirm the look (A5).
3. Call the API. Use the adapter's endpoint and request shape. Prefer rtk curl if available (compact output), else curl. Read the key from the environment at call time (e.g. $OPENAI_API_KEY) — never hardcode, echo, log, or commit a key. Respect the sandbox network allow-list (the adapter's domains must be allow-listed; if a call is blocked, that domain is missing from settings.json).
4. Save the asset to assets/ (create it; git-ignore large/binary outputs unless the user wants them tracked). Use a descriptive filename + a sidecar .json recording {provider, model, prompt, params, time}.
5. Log cost. Append an estimate to assets/COST.md (provider, model, units, est. price) so spend is visible.

Rules

• Keys live in .env (names mirrored in .env.example); they are read at runtime and never surface in output, logs, or commits. The sandbox denyRead + permissions.deny(Read(.env*)) protect them.
• Verify the provider/model/endpoint against the adapter and the provider's official docs (via the sources skill) before calling — don't invent a model name or endpoint.
• Degrade gracefully: if a provider errors or is rate-limited, report it and offer an alternate enabled provider.
• For video/large jobs, respect the seed's cost ceiling / the running assets/COST.md budget. If a job would exceed it, log it as a 🙋 cost-approval item in docs/DEPLOYMENT.md and keep working — don't stop mid-run.