AI Vision

Asks a question about what is on the screen, and answers it in a shape your flow can branch on.

The other detection nodes ask a question you have already answered: is this exact template here (Find Image), what characters are in this box (Read Text), is this rectangle red (Check Region). AI Vision is for the question you cannot phrase that way — which screen am I on, where is the Continue button on a layout I have never seen, is this dialog an error or a reward.

Available in both workspaces.

The answer is constrained, not free text

This is the part that makes it usable in automation. You give the node a Task and a list of Targets; the targets become the only values the answer may take. A model that could reply with any sentence would be useless to a flow — you would be parsing prose at run time. Here, "which of these three screens" can only come back as one of those three.

Task	What you supply	Use it for
Classify the screen	Targets — the screens you expect	"Am I on the shop, the map, or a login wall?"
Locate elements	Targets — the things to find	Finding a control on a layout that moves or that you have no template for
Extract fields	Targets — the field names	Pulling several labelled values out of a panel in one pass
Answer a question	A Question	A one-off judgement: "is the queue empty?"
Describe the scene	—	Diagnostics and Run Evidence: a sentence about what was on screen
Custom (JSON schema)	Your own answer schema	When none of the above shapes fit

Targets, one per line. A target is a short phrase, and a comma inside one is content rather than a separator — the same rule as Open App's arguments.

Settings

Setting	Type	Range	Default	What it does
Task	choice	the six above	Classify the screen	Required. Decides the question asked and the shape of the answer
Targets	list of strings	≤ 60 chars each	—	The states, elements, fields or choices the task is about
Question	string	≤ 300 chars	—	Used by Answer a question and Custom
Region to look at	rectangle	—	whole screen	Where to look. Absent means the whole screen
Timeout	integer	1000–300000 ms	—	How long to wait for an answer

Two expert knobs — the custom task's schema and the image-token budget — are accepted by a saved flow but not offered in the panel. They have sensible defaults, and a hand-authored value is never rewritten by the editor.

Frame it tightly. A region is optional and defaults to the whole screen, which is right for classifying a screen and poor for reading one small panel. Measured on a code challenge: framing the whole card read the heading instead of the code; framing just the picture read the code.

Ports

Port	Taken when	Required
Success	The model produced a schema-valid answer	Yes
Failure	No frame, the runtime was unreachable, or the model could not satisfy the schema	No
Cancelled	The run was stopped mid-question	No — never followed
Fatal	Invalid configuration	No — never followed

Success is not "the answer you hoped for"

"The screen is the shop" and "the screen is not the shop" are both successful reads — the node succeeded at looking. Branch on the content with an If over the bound variables, exactly as you would after a Read Text. A node that routed on the answer's value would need one port per possible state.

Basic example

Know which screen you are on before deciding anything:

• Task: Classify the screen
• Targets: shop, map, login wall
• Save the answer to a variable, then branch with an If on its value.

Common mistakes

Mistake	What happens	Fix
No targets on a task that needs them	The answer has nothing to be constrained to	Fill Targets for Classify / Locate / Extract
Targets that overlap in meaning	The answer is unstable between runs	Make them mutually exclusive
Branching on the port instead of the value	Both answers leave by Success	Branch with an If on the bound variable
Whole-screen region for one small panel	The model answers about the wrong thing	Draw a region around the panel
Using it where a template would do	Slower and less certain than pixels	Find Image for a fixed graphic

Related nodes

• Find Image — when you have the exact pixels
• Read Text — when you need the characters in a box
• If — branch on what came back
• Check Region — colour and change, without a model