What your agent can and can't finish when it makes AI media

What your agent can and can't finish when it makes AI media

An agent gets handed a simple brief: make a short promotional clip for a product launch. It does the hard part well. It writes a prompt, calls an image model, gets a crisp 4K hero frame. It animates that frame into a short clip. It watches the clip back, checks the framing, decides it works.

Then it tries to finish, and finishes nothing.

It wants to smooth the clip and interpolate the frame rate — there's no step for that. It wants to lay a short music bed under the cut — no step for that either. It wants to stamp the output with a "made with AI" credential before anyone sees it — nothing does that as an action it can call. It wants to post the finished file to the launch channel on a schedule — no step. The agent has a good asset sitting in a temp directory and no clean way to ship it.

That is the shape of AI media work for an agent right now: the making is well handled, the finishing is not. Here's which parts of the job you can hand off today, and which parts you'll still have to carry yourself.

What you can hand your agent

The front of the job — turning an idea into an image or a clip, and turning an existing video into something the agent can reason about — is well covered. For each part, here's a skill that covers it.

Make a still from a prompt

To turn text into a picture, nexu-io/nano-banana-one-shop is a runnable path that does the work through a bundled script: it generates a new image, edits one by instruction ("make the sky purple"), and composites up to 14 reference images into a single scene. It exposes an explicit model choice across the Nano Banana / Gemini family — a fast default, a higher-quality 4K option, and a speed-optimized one for high volume — so you can trade cost against quality per call.

If you're targeting GPT Image 2 specifically, evanmopsps/gpt-image-2 plays a narrower role: it drafts a production-ready prompt — subject, style, aspect ratio, resolution, negative constraints — and hands your agent into the GPT Image 2 product, which generates text-to-image and image-to-image up to 4K with legible in-image text (over 95% text accuracy). One thing to know up front: it points out to that product rather than running the generation itself the way the other skills here do. Treat it as a prompt-drafter and handoff — useful when GPT Image 2 is already where you publish, not a local generator.

Make a clip from a prompt or an image

For video, cdeistopened/video-generator drives two providers and tells you when to use which: Google Veo 3.1 when you want native synchronized audio and 4K, OpenAI Sora when you want longer clips (up to 12 seconds) or a particular visual style. It handles the async waiting for you and documents the duration, resolution, and aspect-ratio options for each provider, which is most of what trips up a first attempt.

Put words in a face

When the deliverable is a talking-head, spokesperson, or lip-synced character, prime-skills/ai-avatar-video routes across audio-driven avatar models and picks one for your actual intent — a voiceover UGC clip, a virtual presenter, a dubbed product demo. Feed it a portrait and a voice track and it drives the mouth sync. It also carries the consent and disclosure guidance this kind of work needs, and it will refuse requests that target real people without consent.

Do text, image, and video from one skill

If you'd rather call one thing across modalities, Yacey/agnes-ai-generation spans text, image (including image-to-image), and video — text-to-video, image-to-video, and keyframe interpolation — behind a single interface, with a smoke test to confirm each mode works before you lean on it.

Turn an existing video into something the agent can read

The reverse direction is covered too: an agent can't watch a video directly, so these skills take it apart first. bradautomates/watch downloads a video, extracts scene-aware frames, pulls a timestamped transcript from captions (or a speech model when there are none), and hands the frames and text to the agent to answer questions about what's inside. oxbshw/watch builds on that same shape and adds a persistent index so follow-up questions don't re-process the video, OCR on the frames, offline local transcription that needs no API key, and a capture-critique-fix loop for checking your own generated output against plain-language criteria.

Hand your agent this set and it can generate, edit, animate, voice a face, and analyze — the creative core of the job.

What you'll still have to handle yourself

Past generation, the ground gets thin fast. These are the parts of the job where you shouldn't assume your agent can take the next step — and where, if you're deciding what to build or add, the need is real.

Enhancing a finished video

Stills are easy to enhance; a finished clip is not. There is no skill you can point at an existing video to upscale it, raise its frame rate or smooth its motion, or clean up noise as its own action. The capability isn't entirely missing — one Veo wrapper can bump its own output to 1080p, a style-transfer skill has an optional super-resolution flag, and a routing CLI links out to video upscalers hosted elsewhere — but each of those is welded inside another tool and only reachable if you were already using that tool. Don't assume you can call video enhancement on a clip that came from somewhere else. For now, that's a manual step.

Adding sound to a clip you already have

Every audio skill here is a voice skill — text-to-speech and transcription. None of them generates music, a score, or sound effects. You can get audio if you let a video model bake a synchronized track in at generation time, but if you already have a silent clip you're happy with, nothing scores it after the fact. Your only in-tool path to sound is to regenerate the clip through a model that adds audio during generation — which means giving up the cut you already have. Plan to bring your own music and effects.

Crediting and watermarking the output

If you need to attach a durable "made with AI" credential to what you produced, expect to do it yourself. Some generators can stamp a watermark, but they don't agree with each other — one defaults it on, another off, one brands the mark with its own vendor name, one applies an invisible signal — and every one of them applies only at generation time, to its own output. There's no step that takes an arbitrary asset from a different tool and watermarks it, and nothing attaches a verifiable content credential you can check later. Treat provenance as a step you own.

Publishing the finished file

The last step of the job — actually put the file in front of an audience — is the thinnest of all. Nothing here takes a finished clip and posts it to a platform with a caption and a schedule. The idea shows up only as scraps welded onto other pipelines, never as a step you can call on its own. Assume publishing is yours to wire up.

The short version

The generators are strong. If the job is to make an image, make a clip, put words in a face, or read an existing video back, you can hand it to your agent and expect a good result. The finishing steps — enhance the video, score it, credit it, publish it — are where the work still sits, because there's no clean step for any of them yet. Build for the front of the pipeline today, and plan to carry the back of it yourself.