Eight stages, because one giant request is not a workflow.
Every stage is separate, so every stage can use a different model, and so you can stop at the one that matters and change it.
The pipeline
What actually happens when you press Generate.
Understand the request
Your idea, headline, subject, mood, canvas size and any reference images go to your writing model. It answers with a structured creative brief, validated against a schema before anything else touches it.
Merge the style
A style profile enters the prompt as adjectives, never as content. The strength dial decides how many of its attributes survive; the weakest drop first, the palette holds longest.
Merge persona constraints
A saved persona contributes a written description and up to three reference photographs, so the same face turns up across a whole channel.
Build the image prompt
The brief becomes one flowing paragraph aimed at the specific image model you routed to, densest information first. In overlay mode the artwork is told to carry no lettering at all.
Generate
Your image provider draws it. Capability was checked in stage zero, so a model that cannot draw was never offered.
Typeset the headline
The exact string goes on with HTML Canvas: font, size, weight, alignment, wrapping, stroke, shadow, plate, letter spacing, position.
Save with provenance
Prompt, brief, provider, model, parameters, persona, style, dimensions, parent generation and edit instruction. Never the request headers, so nothing you export can leak a key.
Show it, then argue with it
Click into the editor and start changing things in English.
Before anything is sent
Capabilities are checked, not assumed.
Every model declares what it can do: text, vision, image generation, image editing, masked editing, embeddings, structured output. Re-Prompter reads the provider's own listing where one exists, asks Ollama's own manifest what each local model can do, and falls back to a name heuristic only when neither answers. You can override any of it per model, and your override wins.
The result is that an action you cannot run is greyed out with the reason, instead of a 400 after a thirty-second wait. "This model does not accept images. Choose a vision-capable model, or change your Vision provider."
When something fails
Nothing crosses to another vendor behind your back.
If a provider rate-limits you, Re-Prompter stops and says so, names your configured fallback, and waits for you to choose. Automatic fallback exists, is off by default, and stays off until you turn it on. Quietly re-sending a client's reference photograph to a second vendor is a privacy decision, not an error-handling detail.