What is a workflow and how do I run one?
A workflow is a canvas where you connect prompts, references and models so that one step feeds the next — an image into a video, one reference into several shots — with a cost estimate for the whole graph before you press Run.
What is on the canvas
- Prompt nodes hold text.
- Reference nodes hold images, videos or audio from your Library.
- Model nodes run a generation; their input ports are the model's real parameters.
- Output nodes collect results and keep them.
- Batch and group nodes repeat or organise steps.
Connect a prompt and a reference to a model, the model's output to the next model, and so on. A connection is checked when you draw it, so a wrong port or a missing required input is a red edge, not a paid failure.
Running
Press Estimate to see the credit cost of the whole graph, then Run. Each step waits for the steps it depends on, and up to three run at the same time. A step that fails is not charged, and it stops only the steps that were waiting on it — independent branches carry on to the end. Refreshing the page does not lose a run in progress.
When to use a workflow instead of the Workspace
- You want the same reference reused across several shots.
- One result must feed another model (image → video, video → lip-sync).
- You want to fan one prompt out to several models and compare.
For a single image or clip, the Workspace is faster.
Frequently asked
Do I pay for a workflow before it runs?
No. The estimate is shown first; credits are charged step by step as each generation completes, at the same settled prices as in the Workspace.
Where do the results go?
Into your Library like any other generation, and into the workflow's output node so you can review the whole chain in place.
Can Claude build a workflow for me?
Yes. The connector has create_workflow, wire_workflow, estimate_workflow and run_workflow tools; the assistant hands you a link to review the graph before running.