Custom Flows (Beta)
The Flow Engine lets you replace parts of the built-in generation pipeline with your own node-based flow: a visual graph of prompts, model calls, QA checks, and outputs that runs whenever your character generates content. Open Flows in the sidebar to build one. Available on Pro and Agency plans.
What a flow is
A flow is a directed graph you assemble on a canvas. Nodes have typed input and output ports (a caption port only connects to something that accepts captions), so invalid wiring is caught in the editor, not at run time. Only the part of the graph upstream of the output you request actually executes, and conditional routes skip the branch that isn't taken.
Once published and assigned to a character, the flow runs in place of the built-in pipeline for that character: Fill Queue, Generate Posts, and scheduled generation all route through it.
When to use flows vs the defaults
The built-in pipeline plus character presets covers most rosters. Reach for a flow when you need something the standard knobs can't express: a custom prompt structure, an extra QA gate with your own criteria, best-of-N generation with an automated judge, structured LLM steps, or custom logic between stages. If a default output is merely wrong, tell us, fixing defaults beats everyone maintaining their own workaround.
Node types
- Inputs, read-only data sources: Character, Scene Context, Reference Images, Recent Posts, Brand Goal, Primary Caption, Recent Performance.
- Stages, the built-in pipeline steps as wrappable nodes: Scene Planner, Image Prompt, Image Generation, QA Pass, Caption Prompt, Caption Generation, Video Generation, Reply Chain, First Comment.
- Prompt nodes, raw prompt surfaces: template, custom, and augment nodes where you write the prompt text yourself, with variable interpolation.
- Generator nodes, direct model calls: an LLM node (fixed platform model, with temperature, max tokens, JSON output and schema options), an image model node with a candidates setting for best-of-N, and a video model node. Image and video nodes expose a model picker.
- Judge & QA, a Judge node that ranks candidate images against criteria you write ("sharpest face, product clearly visible") using a vision model and emits the best one, plus a custom QA node for pass/fail checks with retries.
- Code & integration, a Code Transform node that runs your JavaScript in an isolated sandbox (no filesystem, network, or credentials), an HTTP request node, and a webhook notification node. Conditional Route branches the graph on a condition.
- Outputs, terminal nodes that declare what the flow produces: Image, Video, Text. (A Carousel output exists but is authoring-only for now, the editor badges it.)
Flow inputs (surface bindings)
A flow can declare typed input fields. Each field either binds an existing character attribute or publishes a new control onto a platform surface: the character editor's generation card, the creation wizard's visuals step, or the content generate dialog. Surface values are stored per character and reach your prompt templates as {{flowParams.<key>}}, so one shared flow can behave differently per character without editing the graph.
Building & testing
- Dry run, free. Executes the current draft with generator nodes stubbed, so you can verify wiring, prompts, and routing without billing.
- Test run, executes the draft against a real character with real model calls, billed at normal rates. Neither run type creates content items.
- Traces, every run records a per-node trace (status, timing, outputs, errors) you can open from the run history, plus aggregate stats per flow.
- Publish & rollback, editing changes a draft; characters run the last published revision. History lists every published snapshot and lets you roll back.
Attaching a flow to a character
Assign a flow to a character per pipeline kind. Kinds that execute today: the composite content pipeline, caption and text generation, first-comment and reply-chain engagement, and reference-sheet creation. Scene-planner, image-only, and video-only kinds can be authored but not yet assigned (the editor badges them "Authoring only"). If a character has both a composite flow and a caption flow assigned, the composite flow wins, it handles the whole request first.
Pricing & plan availability
The Flow Engine is included on Pro and Agency plans (Beta). Free accounts can see the Flows page but can't create or run flows.
There is no separate flow fee: a flow bills the same per-operation rates as normal generation, images, video, and LLM tokens at the published rates in Billing & Plans. Each best-of-N candidate is billed like a normal generation. Before a billable run starts, a worst-case cost estimate is checked against your monthly budget cap, and blocked runs tell you the estimate.
Related
- Content & Scheduling, the standard pipeline flows replace.
- Brand Goals, available inside flows via the Brand Goal input node.
- AI Assistant, operate the platform conversationally.
- Billing & Plans, rates and budget caps.