> For the complete documentation index, see [llms.txt](https://help.intangible.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://help.intangible.ai/visualize/auto-prompt.md).

# Auto-prompt

How the visualizer assembles the prompt from your 3D scene, camera, and named details.

The visualizer builds the prompt from your 3D scene, your camera, and the named details on your objects. You don't have to write a long string of adjectives – the scene already knows what's in it.

![Visualizer Scene tab with the auto-generated prompt grouped under bracketed headings such as \[Scene Context\], \[Environment\], \[Camera\], \[Subjects\] and \[Lighting\], with the Generate prompt button above it.](https://1179478100-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FSWrjIqlgVAblx6vHGJiM%2Fuploads%2Fgit-blob-a5ba765ce7e4d8415a8b3197f4462d78cce277d7%2Fscene-tab-01.png?alt=media)

## What it does

The default workflow most diffusion tools push you into is: write a paragraph describing your scene, hit generate, see what comes out, edit the paragraph. Intangible flips that. The 3D scene is the spatial truth; the prompt is generated from it. You read the auto-prompt to confirm the model sees what you intended, edit only the parts that are wrong, then generate.

The prompt is structured into three blocks:

* **\[Scene Context]**: the framing – what kind of shot, what mood, what location.
* **\[Environment & Props]**: the physical setting and secondary objects.
* **\[Subjects]**: the hero objects – the car, the character, the product.

Each block draws from a different source. Editing the right source upstream is faster than editing the prompt every iteration.

## How to use it

The auto-prompt lives in the **Scene** tab of the right-hand Visualizer panel.

1. **Open Visualize mode** with a shot active. The auto-prompt appears, populated from the scene and the active shot's camera.
2. **Read the three blocks.** Confirm the model is reading the scene correctly. If a hero object isn't in the \[Subjects] block, the model isn't going to render it well – go fix the source.
3. **Edit a block in place.** Click any block's text and edit directly. Manual edits stick on the next generate.
4. **Regenerate the prompt.** Click **Generate prompt** to rebuild from the current scene state. Useful after you've made Build-mode or Compose-mode changes – the auto-prompt won't update on its own.
5. **Generate.** Pick a model and click **Generate**. The prompt the model receives is the assembled three-block string.

### Where each block comes from

* **\[Scene Context]** is built from the active scene's name and description, the active shot's name and description, the camera lens, and the aspect ratio. Editing a shot's name or description (in [Shot details](https://help.intangible.ai/compose/shot-details.md)) is the cleanest way to influence this block.
* **\[Environment & Props]** is built from populator labels and the names of secondary objects in the scene. Renaming a populator from "populator-1" to "tropical forest" changes this block immediately. Build-mode environment lighting (sun, time, fog) isn't read into the prompt; the one environment input that reaches the render is ground color, which the vision model reads as a material. See [Environment](https://help.intangible.ai/build/environment.md).
* **\[Subjects]** is built from the named hero objects, their descriptions, and any image references attached to them. This block is where attaching an image reference has the strongest effect on the rendered output.

### Objects and characters with references

The auto-prompt used to keep only a few subjects and drop small ones, so a referenced object could miss the prompt and the model put its reference photo wherever it liked. Now every object or character that carries a reference image and is visible in the shot is named, with where it sits in the scene:

* **Each one gets an anchor.** "At the far end of the Pier", "in the Sedan", "behind the Sports Car". The anchor tells the model which object in the frame the reference belongs to; the 3D scene still decides where it goes. People are never used as anchors.
* **Same-kind objects are grouped.** Referenced objects of one kind that you haven't described get one mention, such as "four Market Stalls in a row on the Stone Quay". People, and objects you gave a description, are always named one by one.
* **Characters are described from their pose.** Whether a character stands, leans or lies down, and which way their body and head face, come from the posed figure at the shot's frame, written as one plain phrase. A head turned toward another character names them: "head turned toward Jill".
* **Nothing referenced is skipped.** If the first draft leaves a visible referenced object out, the prompt is sent back once to add it, and a short clause is appended if it is still missing.
* **References outside the frame are left out.** A character just out of shot no longer sends its reference sheet to the model, which would otherwise place it somewhere in the frame.

A shot's stored prompt keeps the wording it was written with. Click **Generate prompt** to rebuild it under these rules.

## Details

| Block                  | Source                                            | Best way to influence                   |
| ---------------------- | ------------------------------------------------- | --------------------------------------- |
| \[Scene Context]       | Scene + shot names, lens, aspect ratio            | Rename the shot; change lens            |
| \[Environment & Props] | Populator labels, secondary object names          | Rename populators                       |
| \[Subjects]            | Hero object names, descriptions, image references | Attach image references to hero objects |

## Limits and known issues

* **The auto-prompt doesn't refresh automatically.** Build-mode changes don't trigger regeneration. Click **Generate prompt** when you've made meaningful scene changes.
* **Manual edits get overwritten on regeneration.** If you've edited a block in place and then click **Generate prompt**, your edits are lost. Keep them, or rebuild from upstream.
* **An anchor can name the wrong object in a busy shot.** Anchors are worked out from what overlaps in the frame and how far away it is, so a crowded frame can pick the wrong larger object. Edit that phrase in the \[Subjects] block.
* **Head direction is the least certain part of a pose.** On an unusual character rig the head can read wrong while the body is right. Only other characters, not objects, are named as what a head turns toward.
* **Long object descriptions get truncated.** The visualizer trims very long descriptions before they reach the model. Keep object descriptions to a sentence or two.

## Further viewing

Phil walks through this concept in depth at the 2026 Production Summit (32 minutes).

{% embed url="<https://www.youtube.com/watch?v=Y3CoibxA_ag>" %}

## Related

* [Scene context](https://help.intangible.ai/visualize/scene-context.md)
* [How the visualizer thinks](https://help.intangible.ai/learn-more/concepts/how-the-visualizer-thinks.md)
* [Image reference](https://help.intangible.ai/learn-more/concepts/image-reference.md)
* [Generate image](https://help.intangible.ai/visualize/generate-image.md)
* [Generate video](https://help.intangible.ai/visualize/generate-video.md)


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://help.intangible.ai/visualize/auto-prompt.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
