Skip to main content
Everything in this section so far describes a workflow as JSON, because that is what the API takes. The workflow editor is the same thing on a canvas — you drag steps, wire them together, and it writes the JSON for you. Open a workflow from the sidebar and it loads straight into the editor.

​
The canvas

Six connected cards on a dotted canvas. brief, a text input. gpt-5.5, marked INFERRED, with a PROMPT port reading a script-writing brief and MAX TOKENS set to 400. seedream-4, with PROMPT and IMAGE INPUT ports and a prompt for the opening frame. seedance-1-pro, with IMAGE, PROMPT and LAST FRAME IMAGE ports, a prompt for the camera move, IMAGE bound to the still step's output URL, DURATION 5, and one more setting collapsed. elevenlabs-tts-turbo-v2.5, with TEXT and VOICE input ports and URL and AUDIO output ports, its TEXT bound to the script step's message content. And Output. Curved wires run brief to gpt-5.5, gpt-5.5 up to seedream-4 and down to elevenlabs-tts-turbo-v2.5, seedream-4 to seedance-1-pro, and gpt-5.5, seedream-4, seedance-1-pro and elevenlabs-tts-turbo-v2.5 each to Output.Six connected cards on a dotted canvas. brief, a text input. gpt-5.5, marked INFERRED, with a PROMPT port reading a script-writing brief and MAX TOKENS set to 400. seedream-4, with PROMPT and IMAGE INPUT ports and a prompt for the opening frame. seedance-1-pro, with IMAGE, PROMPT and LAST FRAME IMAGE ports, a prompt for the camera move, IMAGE bound to the still step's output URL, DURATION 5, and one more setting collapsed. elevenlabs-tts-turbo-v2.5, with TEXT and VOICE input ports and URL and AUDIO output ports, its TEXT bound to the script step's message content. And Output. Curved wires run brief to gpt-5.5, gpt-5.5 up to seedream-4 and down to elevenlabs-tts-turbo-v2.5, seedream-4 to seedance-1-pro, and gpt-5.5, seedream-4, seedance-1-pro and elevenlabs-tts-turbo-v2.5 each to Output.

A brief becomes a script, a still image, an animated clip and a voiceover

Three kinds of card, and the difference matters: The little circles on the sides are ports. Drag from one to another to wire a step’s result into the next step. A wire is not decoration — it is what makes the downstream step wait for the upstream one, and what puts ${steps.…} in reach. The toolbar down the left adds a node, switches between panning and box-select, undoes and redoes, zooms, fits the graph to the screen, and lays it out automatically.

​
Configuring a step

Click a card to open the inspector.
A right-hand panel headed gpt-5.5, MODEL · script, with a Run this step button, a model picker reading GPT-5.5, an empty SYSTEM box, a PROMPT box reading Write a 15-second voiceover script for a quiet product film, Brief, followed by a bound reference to the run's brief input, an add-user-prompt control, and PARAMETERS controls for temperature, max_tokens, top_p, top_k, presence_penalty and frequency_penalty — max_tokens holds 400, the rest marked unset.A right-hand panel headed gpt-5.5, MODEL · script, with a Run this step button, a model picker reading GPT-5.5, an empty SYSTEM box, a PROMPT box reading Write a 15-second voiceover script for a quiet product film, Brief, followed by a bound reference to the run's brief input, an add-user-prompt control, and PARAMETERS controls for temperature, max_tokens, top_p, top_k, presence_penalty and frequency_penalty — max_tokens holds 400, the rest marked unset.

The inspector for a model step, with its prompt bound to a run input

What is worth knowing here:
  • Parameters left unset are not sent. They are not silently defaulted to something on our side — the provider’s own default applies, which is usually what you want.
  • Run this step runs one step using the results the others already produced. If an upstream step has not run yet, the panel says so rather than running the whole workflow behind your back.
  • INFERRED on a card means the chosen model publishes no parameter schema, so the inspector worked the step’s ports out from the model’s known parameters instead of fetching them — and they may be incomplete. It runs the same; the badge is there so you know to double-check the fields.

​
Bindings, on the canvas

A binding like ${input.brief} or ${steps.script.output.choices.0.message.content} is the same binding the JSON uses, and the editor shows it as text because that is what it is — there is no hidden second representation. One form is worth knowing before you meet it:
The | is a fallback: use answer if the step produced one, otherwise results. Handy, and worth using carefully — a fallback means a broken primary reference still resolves, so the workflow keeps running and nothing tells you the first half stopped matching anything.

​
The JSON underneath

A dialog headed Workflow JSON showing the definition's ui.nodes canvas positions for the clip, still, voice and script steps and the brief input, followed by the start of the steps array with the script step's id and type.A dialog headed Workflow JSON showing the definition's ui.nodes canvas positions for the clip, still, voice and script steps and the brief input, followed by the start of the steps array with the script step's id and type.

The same workflow as the API sees it

JSON in the header shows exactly what POST /v1/workflows would store. Use it to learn the format, to copy a workflow into code, or to check what the canvas actually built.

​
Saving, versions and runs

Save publishes a new version. The version dropdown beside the workflow name switches between them, and a run always records the version it executed — so opening an old run shows the definition that produced it, not whatever is published now. Your unsaved edits are kept as a draft, and a draft is yours alone: a colleague opening the same workflow sees the published version, not your half-finished change. Run executes the whole workflow with the values in the input cards. Results appear in the run panel at the bottom, and every run is listed under the workflow in the sidebar.

Scheduling it

Once a workflow works, it can run on a timer — every weekday at 09:00, or once at a time you pick.