DVP Tools / Prompt Engineering

Platform-Ready CinePrompt

CinePrompt belongs in Prepare → Image → Video → Vault. Shape your narrative into a platform-ready prompt, then use an approved image before asking CastLoop for video and save the reviewed recipe to Vault.

Generate prompts Quick start Prompt Methodology

Quick start

Build your prompt in three moves

CinePrompt is a professional bridge. It takes your raw story beats and injects the technical cinematography, lighting, and movement data AI models crave.

Define the Scene

Describe your Subject and Action. Be specific—"A weary astronaut" is better than "A person." Describe what they are doing with physical verbs.

Select the Target

Choose one target in Build, or several target checkboxes in Translate. Each target has its own prompt-format guidance.

Generate and Copy

Review the formatted outputs. Check the subject, composition and output settings before rendering. Model behavior varies; prompt formatting cannot guarantee adherence. Copy and paste directly into your chosen tool.

One app, three modes

Build, Translate, and Stills / JSON

CinePrompt carries three tabs along the top. Your scene description follows you between them, so you can move from a video prompt to a matching still without retyping.

Build Prompt

The core mode — guided cinematography controls assemble a platform-tuned prompt for video models (Kling, Veo, Runway, Luma, Pika, and more). Choose a local preview folder in Chrome or Edge, then press ⚡ Preview Image to render a FLUX 3 Image approximation with your own fal.ai key, whatever target you selected. Each preview is billed by fal.ai and saved automatically to that folder. Your browser may ask you to restore folder access later.

Translate Prompt

Paste an existing prompt from any platform and rewrite it for several targets at once via Claude (bring your own Anthropic key; Anthropic bills the request). Useful when a look works on one engine and you want it everywhere.

Stills / JSON

The former JSON Prompt Builder, now a tab here. Build structured prompts for image models — Midjourney flags, FLUX positive prose or a separate SDXL format with a negative prompt, or JSON creative briefs for Nano Banana Pro / Gemini — with deep photographic controls. Use Use current Build scene → to carry a video scene over, and ↩ Send scene to CinePrompt Build to bring an idea back.

Methodology

Engine Calibration

Build applies deterministic templates locally. Translate and Multi-shot use the shared, source-linked model guidance when asking Claude to rewrite your scene. No model benchmark or output-quality guarantee is implied.

The Platform Filters

Every engine has a different "sweet spot" for description. Our tool adapts your vision to match.

Kling Motion-first, emphasis on physical interaction.
Runway Cinematic terminology and lighting focus.
Midjourney Still-image descriptions, --v 8.2 and optional --ar.
Luma Tonal consistency and environmental bed.
Veo Narrative arc and temporal stability.

Named Looks

Attach visual style data from FrameDNA. This ensures that the generated prompt includes the exact color temperature, grain, and contrast of your reference stills.

Tip

For best results, limit your platform selection to the 2-3 tools you actually plan to use. This allows Claude to focus on more nuanced variations.

Verified 21 October 2026

Model standards and availability

These are prompt targets, not new rendering integrations. Build runs locally without a key. Translate and Multi-shot use your Anthropic key; Preview Image uses your fal.ai key and always renders with FLUX 3 Image. Prices depend on the provider, resolution, duration and references; check the destination service before generating.

Set supported resolution, duration, aspect ratio and reference inputs in the destination tool. Wording such as “8K” does not configure output dimensions. Structured JSON is a creative brief unless the destination explicitly accepts that schema. Attach reference images, video and audio there; a text prompt cannot attach them.

Veo availability: Google's Gemini API Veo 3.1 Preview endpoints shut down on 22 October 2026. This does not mean every provider's Veo endpoint is retired. Gemini Omni Flash 1.1 is available here as a separate prompt target.

Kling 4.0

Kling 4.0 — up to 30 seconds, Omni Reference and multi-keyframe direction; provider availability varies.

Official model guidance

Pika 2.5

Focused short clips, image animation and effects; configure output settings in Pika.

Official model guidance

Runway Gen-4.5

Text-to-video and image-to-video with camera choreography; 2–10 second clips on Runway web.

Official model guidance

FLUX 3 Image

Still generation, reference editing and optional bounding-box layout control.

Official model guidance

GPT Image 2.5 Sunburst

OpenAI — precise image generation and editing; choose Flare for faster iteration.

Official model guidance

Nano Banana Pro

Gemini 3 Pro Image — premium image generation and editing; not a video model.

Official model guidance

Seedance 2.5

ByteDance — up to 30-second audiovisual sequences, multimodal references and timed edits.

Official model guidance

Midjourney V8.2

V8.2 still-image composition and aesthetics; Edit Model replaces older character-reference features.

Official model guidance

Veo 3.1

Native audio; Gemini API Preview endpoints close 22 October 2026. Check your provider.

Official model guidance

Luma Ray 3.2

Multi-keyframe control, HDR and EXR workflows; duration and export support vary by operation.

Official model guidance

MiniMax H3 (Hailuo 3.0)

Multimodal video with native stereo audio; 4–15 seconds, with a separate 2K regeneration stage.

Official model guidance

Grok Imagine Video 1.5

Video with dialogue, ambience and motion; choose Grok Imagine Image 2.0 for stills.

Official model guidance

Wan 3.0

Up to 30-second audiovisual clips and multimodal references; limits vary by region and endpoint.

Official model guidance

Gemini Omni Flash 1.1

Google — conversational video generation, editing and extension; successor path for Gemini API Veo previews.

Official model guidance

Nano Banana 2.1

Google — efficient image generation and editing, multi-image references and search grounding.

Official model guidance

GPT Image 2.5 Flare

OpenAI — faster image generation and editing for iteration; Sunburst prioritizes precision.

Official model guidance

Grok Imagine Image 2.0

Image generation and editing with up to five reference images; separate from Imagine Video.

Official model guidance

Seedream 5.0 Pro

ByteDance — multilingual layouts, dense text, reference editing and photographic stills.

Official model guidance

Luma Uni-1.1

Reference-guided images and natural-language editing; separate from Luma Ray video.

Official model guidance

FLUX.2 Max

High-quality stills and reference editing with optional grounding search.

Official model guidance

FLUX.2 Flex

Image generation with adjustable steps and guidance; typography and fine details.

Official model guidance

FLUX.2 Pro

Retained FLUX.2 Pro image generation and editing; FLUX 3 Image is the current preview model.

Official model guidance

FLUX 3 Image Klein

Compact open-weights still model with reference editing and bounding-box controls.

Official model guidance

FLUX 3 Video

Video with synchronized sound, multi-shot direction and up to 20-second clips.

Official model guidance

Qwen-Image-2.1

Unified image generation and editing, transparency and up to ten reference images.

Official model guidance

Krea 2

Aesthetic exploration, style references and moodboards for still images.

Official model guidance

Controls

Mastering the Input

Subject & Action

The "Hero" data. Describe clothing textures, facial expressions, and specific physical maneuvers to anchor the AI's attention.

Environment & Time

Set the stage. "Overcast afternoon in a brutalist library" gives clearer setting and lighting direction than "In a library."

Camera Movement

Describe the lens and the move. Use Dolly Zoom, Handheld Shake, or Slow Push-in to guide the model's temporal engine.

Mood & Style

Emotional keywords. Use "Melancholic noir" or "Tense documentary" to steer the color and contrast logic of the generated output.

Prompt History

Recall your last 20 generations from browser storage. Iteration is faster when you can return to a previous baseline with one click.

A/B Compare

Tick Compare on two history entries and run A/B Compare to read them side by side. Use it to settle which phrasing actually earned the better shot instead of guessing from memory.

Live Preview

See the assembled prompt update as you change the dropdowns, before you spend a generation. What you read in the preview is exactly what gets sent.

Check Prompt

Scores the finished prompt locally and lists each weakness with a concrete fix, then names the model it would route the shot to and the risks that come with it. No API call, so run it as often as you like.

Save to Vault

Send the generated prompt to DVP Vault as a browser-local recipe. The platform, original builder input, camera body, lens, aperture, shot type, movement, speed, lighting, grade, genre, time period, and aspect ratio travel with it. Allow the Vault popup to complete the same-origin handoff.

Load from Vault

Pick a saved recipe and the prompt, original subject, platform, and supported builder controls repopulate. The picker closes after the handoff, so you can continue refining immediately instead of rebuilding a look from a pasted string.

Regenerate

Hit Regenerate to re-run the builder against your current fields — the same deterministic pass Generate runs, with no API call. Change a field first if you want a different result; it won't vary the output on its own.

Vault status: a suite save is a draft in this browser. Review it in Vault, add any ownership or rights notes it needs, then export Vault JSON if it should be promoted to the canonical library.

Professional Workflow

Shot Development

Use CinePrompt during Prepare. It can describe an image or video, but it does not choose a hidden first frame: approve the image in CastLoop before the Video request. Vault saves the recipe and lineage, not a durable CastLoop media attachment.

The DNA Integration

Scan a reference still in FrameDNA, save the look, and apply it here. It's the fastest way to maintain visual style across different prompt descriptions.

Motion Anchoring

If your action is complex, put it at the very start of the Kling prompt. CinePrompt does this automatically for most configurations.

Shot Type vs Camera

Describe the SHOT (Close-up, Wide) rather than just the CAMERA move. AI models respond better to framing terminology than spatial coordinates.

Multi-shot packs

Character-locked prompt packs across platforms now live on this page at #multishot. The LoRA lab that used to sit beside them is identity work and lives in Cast.