Build flows with visible, reusable inputs: connect Text sources to named node fields, bind live values from the linked project Brand, and use typed media/reference sockets for assets. Avoid hidden inline prose and copied Brand content.
Where: Flowgen canvas → Add a source node, select a destination node, and wire source outputs to its named input fields.
Before you start
- Link the intended Brain Brand to the project before relying on its identity fields.
- Use describe_flow or the canvas to confirm node IDs and input handles before wiring.
Key ideas
- A prompt is a connected source: The prompt belongs in a Text input node connected to the generation node’s prompt field, not in miscellaneous node props or unconnected card copy. The create-node prompt helper materializes the connected source for you; connect_nodes lets you inspect or change the edge explicitly.
- Brand data is structured and live: A linked Brand includes typed properties and tagged references. Bind the exact populated compatible field to a node input when a visible Brand chip or stable live reference is useful; later Brand edits then flow through without copying stale prose.
- Use the input type that matches the source: Text, image, video, design-system, reference, and other sockets have different meanings. A reference image belongs on a reference input; an image meant for editing belongs on the image input; a Brand property belongs on its compatible typed field.
- Keep provenance visible: Named wires show where a value came from and make a flow easier to review, change, run selectively, or hand to another person.
Steps
- Open the project’s Brand setting and link the specific Brand record that should ground this project. If none is linked, ask the owner to link or create one instead of inventing brand copy.
- Add the operation node (for example, image generation, video generation, Designer, or a Transform). Give it a useful name so its purpose is clear on the canvas.
- For a new prompt, add a Text source node or use the node-creation prompt helper, which creates a real connected Text input. Keep prose out of unrelated node props.
- Select the destination field you mean to fill—such as Prompt, System Prompt, Persona, reference image, first frame, or design system—and connect the correct typed source output to that field.
- When the desired value already lives in the linked Brand, inspect that Brand record and its structured fields. Use connect_brand_field without field_key first to see compatible fields, then bind the exact returned field key to the exact destination handle.
- For assets attached to a Brand, choose the relevant tagged reference or asset and connect it to the node’s media/reference input. Do not turn an image or URL reference into copied prompt prose when a typed asset input exists.
- Read the canvas and destination sidebar again. Confirm the source chip appears in the intended field, the noodle lands on that field (not merely the card), and multiple connections appear in the intended order.
- Run only after the source content, model, target, and any Spark cost are reviewed and the user has authorized execution.
A safe MCP sequence
- Call describe_flow/get_flow_state to use real IDs and inspect the project’s linked_brand context.
- Call read_record on the Brand when you need an inner property; use its brandProperties result for canonical keys, labels, types, exact values, and usage.
- Call connect_brand_field with node_id + target_handle and omit field_key to list compatible Brand fields without changing the canvas.
- Bind the selected exact key, then call describe_flow again to confirm the live pointer and target field.
- For ordinary user-authored prompt text, use create_node(prompt) or set_prompt so the value is represented by a connected Text source; do not place prose into arbitrary set_props fields.
What should be wired where
- A written brief or instruction → Text source → prompt_in or system_prompt_in.
- The linked Brand’s tone, audience, mission, or guideline → that Brand’s exact compatible structured field → matching text/persona field.
- A style or product image → reference_in when it should guide appearance; image_in when the operation should edit that image.
- A project design system → design_system_in, not a prose prompt or a similarly named generic content input.
- A video clip → video_in; a first frame, last frame, and style reference use their distinct roles on a video-generation node.
Lists and repeated work
- Use Items/List plus List Iterator to process rows and a Results List to collect typed outputs.
- Group Repeater, loop-enabled Group, and Batch are deprecated; keep them only to inspect or repair existing canvases.
- A plain Group remains useful for organization and typed exposed inputs/outputs, but it is not the new repetition mechanism.
Tips
- Name source nodes (“Campaign brief”, “Product reference”, “Brand voice”) so the graph explains itself.
- If the correct Brand field is not offered, check the linked Brand, its populated values, and the destination socket type before trying another field.
- Prefer live Brand bindings for reusable identity rules; use a Text node for a brief that is specific to this flow.
Limitations and important notes
- Only fields compatible with the target handle can be bound; similar labels do not guarantee compatible types.
- A Brand reference and a typed Brand field are not interchangeable: references preserve source assets/context, while fields provide structured values.
- A direct prompt source is still user-authored content; creating the node does not grant permission to run a paid operation.
Troubleshooting
The Brand field picker is empty.
Confirm the project has the intended Brand linked, inspect that record’s populated structured fields, and confirm the destination handle accepts that field type.
The wire attaches to the wrong field on the card.
Use the named socket/field as the destination, not the card body. In MCP, pass the explicit target_handle after inspecting the node’s actual handles.
A node is still using old copied brand text.
Replace the copied prompt fragment with the correct live Brand field binding where compatible, or retain a Text source only for flow-specific wording. Then verify the resulting source chip and connection.