Stages AI CUE rewrites text-to-video prompts before generation
Dustin Hollywood posted Stages AI CUE text-to-video tests where assisted prompting reshaped creator directions before generation. He tied the demos to continuity checks, reusable casting tools, and long-context prompting.

TL;DR
- CUE is an assisted prompt rewrite layer, with dustinhollywood's Dragon Rider post saying it takes creator directions and prompt, then reshapes them before generation.
- The continuity recipe is long-context text-to-video plus structured descriptors, with dustinhollywood's continuity reply describing continuity checks, agent QA checks, and reusable casting tools.
- EDITX adds a model-switching motion-editing surface, with dustinhollywood's EDITX post listing nine motion editing models, saved presets, LUT/color work, reframe, restyling, and upscaling.
- STAGES is also pushing creator cost accounting, with dustinhollywood's cost-tracking post showing per-generation spend tracking, model economics, burn rates, and exportable reports.
- The platform is BYOK-friendly at the creator tier, with dustinhollywood's INKROOM post naming a $29 monthly plan and support for frontier, BYOK, and open-source model workflows.
The official STAGES site says CUE can take a thought, expand it into a detailed brief, and render from the live engine. Its About page frames STAGES as one command surface for agents, generation, asset management, review, and delivery. A STAGES YouTube session shows CUE stretching rough fashion-video direction into longer cinematic prompts, while a July 10 LinkedIn post says the platform also added 102 six-second VFX transition presets.
CUE-assisted prompting
Dustin Hollywood, listed on the STAGES About page as founder and CEO/CCO, framed CUE as an assisted layer with the creator still in the loop. The post says CUE takes a user's directions and prompt, rewrites them, then sends the result into generation.
The workflow is small and important:
- The creator supplies direction.
- CUE expands it into a more detailed technical prompt.
- The creator generates after the rewrite.
Hollywood wrote that AI-assisted prompting is “10x better than autonomous” because an agent that misses details creates another iteration loop. dustinhollywood's second Dragon Rider post made the same argument with another capture of the prompt expansion flow.
Long-context continuity
Hollywood said the examples in this batch were text-to-video, and argued that T2V can hold strong consistency when the prompt carries enough context. The Man on a Train clip keeps the subject's face and clothing steady across a slow shot by a train window.
His follow-up described the continuity system behind the demos, according to dustinhollywood's continuity reply:
- Start with characteristics that are abundant in model training data.
- Stagger descriptors from general traits to finite detail.
- Run continuity checks.
- Run agent quality checks.
- Store characters, locations, and objects in a casting tool connected across the platform.
- Reuse those assets by opening the drawer, clicking them, tagging them, and generating.
The test set covered a few different failure modes: a stable seated subject in the train clip, a portal transition in the Escape clip, a mechanical character study in the Father Time clip, and a vehicle crash sequence in the Crashing On Paradise clip.
EDITX motion editing
EDITX is presented as the place to alter existing clips with model choice, presets, color, and motion controls in one surface.
- Model surface: nine motion editing models, per dustinhollywood's EDITX post.
- Inputs: load a clip from device or STAGES assets, per dustinhollywood's load-clip post.
- Controls: saved presets, LUTs, color editing, tones, composition, motion, depth, reframe, restyling, and upscaling, per the EDITX post.
- Restyling models: HappyHorse, Wan, Kling, Grok, Luma 3.2, and LTX, per dustinhollywood's restyle model note.
- Partial-change models: Aleph 2, HappyHorse, Grok, Seedance, Veo, and LTX, per dustinhollywood's partial-change note.
- Alternate surface: those models can also be used in VIDEO Gen without EDITX's saved-preset extras, per the same partial-change note.
RESOLVE Audio Studio
The RESOLVE demo shows STAGES treating generated audio as editable timeline material.
Hollywood's list of tools was direct:
- Drag recent generations onto tracks.
- Record live on any track.
- Apply FX, including Echo, Tremolo, Rumble, Drive, and Limiter.
- Extend, split, and duplicate tracks.
- Extract stems.
The same evidence item includes a separate enhancement note: Hollywood said a 4K sample used Proteus plus Rhea on layered enhance, then argued that model choice should follow the task rather than the newest release badge.
INKROOM and BYOK
INKROOM is the drawing side of the STAGES surface: hand drawing, digital illustration, image layering, inpaint, gen fill, and a visible iteration history timeline.
The tool list in dustinhollywood's INKROOM post includes:
- Ballpoint pen.
- Ink pen.
- Fine line art.
- Highlighter or marker.
- Eraser.
- Collage image layering.
- Inpaint and gen fill.
- Iteration history.
Hollywood named GPT-Image 2, Nano Banana 2, Flux 2 Pro, and more as available choices. The same post says STAGES supports BYOK, and the official pricing block lists a $29 Creator plan with BYOK routing.
Production costs
Hollywood framed cost tracking as a core product choice: dustinhollywood's cost-tracking post says every generation and every project has real production costs attached.
The accounting surfaces listed there:
- Cost tracking for every generation.
- Real production cost per project.
- Model economics before generation.
- Output comparison.
- Burn tracking over time.
- Exportable reports.
The screenshots show spend split by modality, top models by spend, live credit events, rollover status, and a 234-page credit spend report.
Storyboard timelines
A separate creator, DavidmComfort, posted a WIP AI video platform and a Boston Tea Party documentary during the same window. The screenshot is useful because it shows the same production grammar showing up outside STAGES: scene timeline, cast sheets, blocking board, setting plate, and render settings in one board.
The documentary post runs nearly three minutes and uses historical illustrations, a digital narrator, colonial ships, and title cards for the Boston Tea Party.
Pre-vis at 480p
Hollywood's most wallet-aware workflow note came from a reply: pre-vis at 480p, upscale to 1080p, then upscale again to 2K. He called 4K a novelty unless the work is headed to a screen that needs it.
The attached comparison claimed the 480p source was generated faster at about a quarter of the price.