The most useful answer is sometimes no image
Automatic illustration attaches decorative photographs to every item. The technical explainer gets a photograph that adds weight without adding meaning. The one page that actually needs a diagram gets the same treatment.
The real work is choosing the type. A process should have a diagram. A comparison should have a chart. A human story should have a photograph. If a concept has no spatial structure, you usually need nothing. Forcing an image onto it makes the page worse.
The prompt matters only after the type is chosen. A prompt written after the type is selected is a different kind of prompt. It knows what the image needs to do, not only what the page is about.
If you force the agent to give a reason, such as "nothing here needs a picture", the output becomes something you can reject.
How the agent is built in Agent Studio
The system prompt lists your brand's visual types and what each is for. It also states that producing nothing is a valid outcome and must be justified the same way as any other.
Prompt skills hold the house style block: the palette, the treatment, and the things your brand never shows. When the brand is refreshed, every later prompt updates with it.
The agent reads the content, decides, and calls Generate Image only when an image is necessary. When it decides no image is needed, it gives the reason and produces no asset.
In an ActionFlow, the Agent Chat node puts this decision into a content production process. The flow produces the copy, asks the agent what kind of visual is required, and only then calls Generate Image with the prompt the agent wrote. The image is guided by the content, not chosen alongside it.
Before anything is made public, a person reviews it, as with every generated visual.
Cost is shown per node, so the decisions not to generate an image become visible in a way they usually are not. Over one month, you can see what the agent decided not to spend.
What this agent is built from
- System prompt: your visual types, what each is for, and no image as a justified outcome.
- Prompt skills: the house style block injected into every generation prompt.
- Generate Image (AI Core): called only when the agent judges an image necessary.
- Agent Chat (Util): places the decision inside a content production flow.
- Slack (Human in the Loop): review before anything generated ships publicly.
- Node cost outputs: make the value of no-image decisions visible over time.
- When an ActionFlow is enough: If you need templated assets such as quote cards and release banners, build a workflow. This agent is for editorial work where the correct visual varies each time.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.