Tasks and prompts

Every /design call is driven by a task — a textual instruction describing what Rivermark must do. It's interpreted as must-follow instructions, like a brief you'd hand a trusted designer.

There is no special syntax to learn — describe the job the way you'd describe it to a person:

• the deliverable and its shape — "create 3 Instagram story options", "resize to LinkedIn format",
• hard requirements — "use only the fonts from the brand kit", "no AI-generated imagery",
• copy to typeset — "apply this headline and subline exactly",
• how references map to the job — "use the first reference as the layout and the second as the brand".

Examples

Create with copy:

Create an Instagram post from the provided brand kit with the following text:
"Summer sale — 30% off everything. This weekend only."

Resize with reflow:

Resize the provided design to Instagram story (1080x1920) and LinkedIn post
(1200x627) sizes. Reflow as needed rather than cropping.

Work from a template:

Based on the provided carousel template, create a short explainer carousel
for the referenced blog post. 5 slides maximum.

Generate ideas:

Create 6 ad creative ideas for the user's brand based on the provided
best-selling product list.

Prompting guidelines

Be specific about outputs. "Create a poster" is fine; "Create a 1080x1350 poster announcing the event in the reference" is better. Even better, declare the outputs explicitly (see Outputs).

Name your references. With several references, tell the agent which is which: "the first reference is the brand kit, the second is the product photo to feature."

State constraints once. Anything in the task is enforced; anything omitted is left to the agent's judgment (moderated by the creativity level). If something must never happen — e.g. never invent copy — say so.

Don't prompt-engineer around quality issues. Overflows, clipping, legibility, and invented content are caught by auto-review, not by prompt rules. If you see a defect class that survives review, tell us.

When a human is behind the screen

If a real person's words are relayed to the API — an end user in your editor, or a teammate — pass them in the user_prompt field instead of inlining them into task. user_prompt is interpreted more freely than task: it handles ambiguities in the prompt better — users say things like "make it pop" or "less corporate" — and provides a smoother editing experience, while your task stays in control of scope.

{
  "task": "Apply the user's requested changes to the design. Do not change anything else.",
  "user_prompt": "can the headline be smaller? also more blue",
  "continue_from": "rmd_cj4hbqz2w0003mz6x7p9q2xe"
}

This split is what the AI design editor guide builds on.