Skip to main content
Back to Blogguides

The Content Agent Contract For Manual AI Video Workflows

The content agent should not pretend it has direct image or video generation APIs. Its job is a committed HTML package with prompts, hooks, captions, and publishing notes.

OnyxWork Content Systems|2026-06-11|8 min read

Some content agents fail because they pretend every creative tool has an API. Others fail in a quieter way: they produce a nice-looking document that cannot actually be used by the person making the asset. The useful contract is narrower and stricter. Research what should be made, define the hook, write the exact prompts, package them in a durable file, commit it, and notify the human who runs the media tools.

For OnyxWork, that means the content agent does not directly generate final videos. It creates the production package for manual OpenArt, GPT Image 2, and Seedance 2.0 workflows.

The Deliverable Is A Repo Artifact

The daily output is a branded HTML file at marketing/daily-viral-content-YYYY-MM-DD.html. It must be committed to the platform repository. A file saved in a vault, a downloads folder, or a container scratch path is not complete. The founder needs a GitHub URL, a concise summary, and a file that can be opened in a browser while producing assets.

This is why HTML works well. It keeps prompts copyable, visual sections separated, status notes visible, and the daily package trackable in Git. A chat transcript is too easy to lose. A markdown note is fine for memory, but not enough for production. The HTML file is the working surface.

Reference Images Come Before Video

AI video models are good at motion and weak at stable invented interfaces. If a prompt says animate a kanban dashboard, the model may produce warped cards, unreadable labels, drifting typography, or fake UI. That is not a Seedance problem alone. It is a pipeline problem.

The fix is reference-first production. Any visual that contains UI, charts, dashboards, terminals, social posts, product screens, logos, or text must be generated as a reference image first. The video prompt then tells Seedance to animate only those references. It should not ask Seedance to invent new UI from prose.

Every Segment Needs Its Own Prompt Logic

Long video prompts become sloppy when they try to describe an entire piece at once. The daily package should split each concept into chained ten-second segments. Each segment needs its own reference asset plan, image prompts, motion instructions, and final Seedance prompt.

If a segment lists three reference assets, it needs three distinct image prompts. Not one prompt with vague instructions to make the rest. Not a storyboard note. Not a summary. The asset plan and prompt count have to match.

The Right Section Shape

A usable segment has a predictable anatomy:

  • Reference Asset Plan: numbered assets and the exact order to feed into Seedance.
  • GPT Image 2 Reference Prompt N: one copy-paste prompt for each required image, written so it also works in OpenArt.
  • OpenArt Settings Notes: aspect ratio, style, negative prompt, reference-lock advice, and text spelling warnings.
  • Motion Instructions: exact timing, camera movement, transition logic, and what must stay locked.
  • Seedance 2.0 Final Prompt: the actual paste-ready video prompt for that segment.

What The Agent Must Avoid

The content agent should not write agency theater. Terms like creative blueprint, director's read, and production memo can make a file look polished while hiding the fact that the prompts are unusable. The founder does not need a film-school document. The founder needs copy-paste-ready inputs that make the models behave.

The agent also has to avoid unverifiable claims. If it says a trend is current, it should have source evidence. If it shows a product dashboard, the dashboard should be defined in a reference image prompt. If it mentions OnyxWork, it should use onyxwork.dev and no hallucinated domain.

SEO And Daily Viral Are Different Jobs

The same content agent can support SEO and short-form media, but the deliverables are different. SEO work should update real website content paths, metadata, internal links, and blog drafts. Daily viral work should create prompt packages for manual media generation. Mixing those contracts is how agents end up saving a random HTML file locally and calling it published.

A senior-grade content agent knows the difference between a draft, a staged asset, a committed package, and a live website page. It should not claim publication until the correct repo path changed and the resulting page or file can be inspected.

The Standard

The daily package is complete when it contains researched hooks, eight video concepts, strict ten-second segments, matching reference image prompts, OpenArt-compatible settings, Seedance final prompts, captions, publishing notes, and a GitHub URL. Anything less is not a content pipeline. It is a mood board with extra steps.

Written by

OnyxWork Content Systems

Growth Systems

OnyxWork operators documenting the systems behind governed autonomous work.

Production context, not generic commentary.