# How do you track prompts, references and versions for AI-generated media?

Track each durable creative object as an Asset with sibling Variants, not as a folder of loosely named files. For every generated Variant, retain its typed Recipe, prompt, settings, and exact ordered input References; then explicitly select the main Variant that currently represents the Asset. This record works across images, video, and audio, and it preserves both unsuccessful directions and the chosen handoff without pretending that the same inputs guarantee identical media.

## Why is saving the prompt alone insufficient?

A prompt omits the exact inputs, typed settings, operation, output, and selection. A complete record answers: what Asset was made; which Variant resulted; what its Recipe requested; which exact Variants were References; and which Variant is main. [Core Concepts](/docs/concepts) defines each term.

## What does a complete creative record look like?

The public [Night Market recorded session](/#showcase) is backed by shipped replay data. Its ten Assets include images, audio, and a derived video:

| Record | Visible evidence |
|---|---|
| Variants and selection | Mara has three completed Variants; `4904331f` is main. |
| Recipe and References | The vignette stores its prompt and three ordered exact References. |
| Computed Reference graph | Three Recipe Reference edges lead to the vignette; Mara also leads to the video. |
| Cross-media work | The rain effect and theme retain audio Recipes; the idle loop retains a video Recipe. |

![Composite with three exact References in its stored Recipe](/showcase/8de24c27-38eb-4ab3-a59a-266751215cda-inspect.webp)

## How do you capture the record during production?

Generate an Asset; inspect the completed Variant's Recipe and References; Regenerate unchanged replays or Vary deliberate changes as siblings; Retry only an unchanged failed Variant; compare the media and Activate the selected result; then use exact Variant references for dependent Assets. The Recipe is the immutable receipt, and its ordered References compute the input graph.

The recorded video used `asset:mara-the-lantern-keeper~0f4af497@4904331f` and returned `asset:mara-idle-loop~7d22189e@3828d140`. Those compact references connect two distinct Recipes and results.

## Does this make generation deterministic?

No. The record makes work inspectable and reusable; it does not guarantee identical pixels, frames, motion, or audio. Keep the exact completed media as evidence and compare new siblings against it.

Use the [CLI guide](/docs/cli) or [MCP guide](/docs/mcp) to inspect the shared record, and the [audio playbook](https://makefx.app/docs/audio-playbook.md) for medium-specific work. Then read [how Variant selection works](/docs/ai-media-variants), or [follow the quickstart](/docs/quickstart) to create the first tracked Asset in a Space.