Ocean Ring

An explorer walks toward an ocean-spanning ring, with waterfalls beneath its glowing rim.

Download HDR SDR

Showing SDR, which plays on any screen.

Vivid HDR

Ten-second worlds, generated by AI and graded so their highlights can shine brighter than the white of your screen. Here's how they were made, and a prompt that teaches your own coding agent to make more.

Inspired by "grassland special", an HDR video by @hi_personality, with instructions adapted from @EianLu's breakdown of how it gets its brightness.

Get the prompt

Three steps, in this order

Start with a bright world

Generate one detailed first frame with luminous whites, turquoise water, vivid greens and warm orange accents. Use full sun and plenty of fill light so the midtones stay colorful. HDR can't rescue a frame that was composed dark.
Make it really move

Send that frame to a video model with a specific action and camera path. A skater pushes and carves, a walker's feet alternate, and a drone follows close enough for the foreground to rush past. A slow zoom over a still doesn't count.
Give the highlights room

Convert the footage to linear light, move it into BT.2020, push only the brightest parts upward, and encode 10-bit HLG in HEVC. Keep a separate SDR export for screens that can't show the difference.

Colorful isn't the same as HDR. These scenes should look vivid in SDR too. HDR adds something else: on a compatible screen, clouds and sun glints can go brighter than the white around the video. Screenshots and GIFs can't capture that, so judge it on the screen itself, ideally with the brightness turned down.

Sharing on X? Post the HDR file. That's the version that can outshine the feed around it. You don't generate anything twice: the skill makes one clip and exports both files from it. Keep the SDR file for places that don't handle HDR well, and for comparing the two. X re-encodes uploads and we haven't tested how our clips hold up there, so check the post on an HDR phone.

What you'll need

Your coding agent runs the workflow. The cloud renders the video and your computer does the final encoding, so you don't need a powerful graphics card.

A coding agent with file access, a terminal and internet access, plus support for reusable skills or saved instructions.
An OpenRouter account with credits and an API key, kept in a private environment variable and never pasted into a prompt or page.
A first frame. Supply your own, or let the agent make one. Codex uses its built-in image tool. Claude Code has none, so it uses OpenAI's latest image model through OpenRouter.
Local finishing tools: Python 3, NumPy, Pillow, and FFmpeg with libx265 and lut3d. The agent can check for these and walk you through installing them.
An HDR screen and player to see the result. A recent iPhone or Mac in Safari is the safest bet.

One prompt to build the skill

A skill is a folder of instructions and helper scripts your agent can reuse. Paste this into a coding agent. It installs the workflow and tests it locally without spending anything on video.

Paste this whole prompt into your agent

Create and install a reusable skill named vivid-hdr-video for this agent. It should take a scene prompt or reference image and produce one genuinely animated, brilliantly colored fantasy or science-fiction video, plus verified HDR and SDR exports. Build the skill and supporting scripts now. Do not start a paid generation just to create the skill.

STYLE
Aim for a sunlit, richly detailed fantasy-game cinematic: electric-cyan water and sky, luminous white architecture, emerald foliage, warm orange accents, colorful midtones and generous ambient fill. Adapt the palette to the subject. Avoid gloomy exposure, muddy shadows and a global bloom that erases detail. Make a bright starting image first, or use my supplied image. If this agent has its own image generation tool, as Codex does, use that tool. If it doesn't, as with Claude Code, use the newest OpenAI image model on OpenRouter, in its precision tier when a release has several (openai/gpt-image-2.5-sunburst as of November 2026). Find it in the live catalog at https://openrouter.ai/api/v1/models?output_modalities=image and call POST /api/v1/images with an aspect_ratio such as 16:9 and high quality; that cost about $0.03 per image in November 2026. Inspect the image before making a video.

MOTION
Generate actual subject motion and a moving camera, not a pan or zoom over a still. Describe a coherent ten-second action and a drone-like tracking path, with foreground parallax and environmental motion. Walking needs alternating footfalls and weight shifts; riding needs propulsion, turns and believable speed changes. Keep anatomy, architecture, exposure and palette stable. Default to one silent 10-second, landscape 1080p clip; honor explicit counts and settings.

PROVIDER AND COST
Use OpenRouter's video-generation API when available. Read its current official documentation and discover supported models, settings and prices at runtime. Prefer kwaivgi/kling-v3.0-pro while it is offered, and explain any other choice. It cost $1.12 for a silent 10-second 1080p clip on November 2, 2026, when the API accepted 1080p and returned 1920x1080 even though the catalog's supported_resolutions listed only 720p; do not assume this price or availability still applies. Send a local first frame in frame_images as a base64 data URL with frame_type first_frame, which worked in that same run. The API can't list past jobs, so reconcile a job with no saved ID through the OpenRouter activity page. Read OPENROUTER_API_KEY privately from the environment. Explain the proposed model, upload and estimated batch cost before spending; honor existing authorization within its scope. Never print or store keys, or put them in browser code.

Implement submission, polling and downloading with durable job records. Reserve the output folder before a paid POST and save the job ID immediately. Do not blindly resubmit after a timeout or charge for extra variants. Reconcile uncertain jobs first. Handle failed, cancelled and expired jobs, timeouts and safe HTTPS redirects; never forward authorization to a different host. Preserve sources, prompts, settings, job status and returned cost without secrets or signed download URLs.

REAL HDR CONVERSION
Use Python 3 with NumPy and Pillow, and FFmpeg/ffprobe with libx265 and lut3d (or an equivalent verified color-management pipeline). Check dependencies and explain how to install missing ones. No local generative-video GPU is required when the provider renders the video.

Inspect the source's color tags. For known BT.709 SDR, decode its transfer curve to linear light, convert BT.709 primaries to BT.2020, perform controlled highlight expansion, then encode HLG with a documented reference display/OOTF assumption. Do not merely relabel SDR as HDR. Do not apply the conversion twice to an already-HDR source. Require an explicit, recorded assumption for untagged sources.

Preserve midtones and readable shadows; gently enrich saturation. Give clouds, reflections, sunlight and luminous structures extra headroom, with a soft highlight rolloff. Treat highlight-area percentages as creative targets, not proof of physical display brightness; our clips landed at 10-18% of luma above 75% and 3-9% above 90%. Scale ordinary SDR tones consistently to HDR reference white (approximately 75% HLG under the chosen reference conditions, 203 nits on a 1000-nit display), then expand only the brightest part of the scene, roughly the top 18% by luminance, up to about 950 nits. SDR white therefore lands above 75% on purpose. Avoid clipping; disclose encoding overshoot. Export HEVC/H.265 10-bit yuv420p10le, HLG/arib-std-b67, BT.2020 primaries and bt2020nc matrix, MP4 tagged hvc1 with faststart. Preserve audio when present. Also export a browser-friendly H.264 SDR comparison. An optional GIF is SDR only.

VERIFY AND DELIVER
Use ffprobe to confirm codec, bit depth, transfer, primaries, matrix, duration and frames. Decode the complete output to check errors. Read raw 10-bit Y directly, normalize using the file's declared range (limited-range nominal codes 64-940), and measure above 75%, above 90%, and out-of-range values across frames. Do not use auto-stretched grayscale. Create a diagnostic map with yellow above 75% and orange above 90%, plus a contact sheet. These are encoded-luma measurements, not nits or proof of HDR playback on a phone.

Inspect the moving video for real travel, camera perspective changes, stable subject geometry and consistent brightness. Explain any visible defects. Deliver the HDR MP4, SDR MP4, validation JSON, grading settings and a hash manifest, with distinct filenames. Tell me to post the HDR MP4 when sharing on X or similar feeds, and to keep the SDR MP4 for places that don't handle HDR well; both come from one generated clip. Do not claim iPhone or social-platform HDR playback has been tested unless it has. Actual brightness depends on the screen, player, OS and platform transcoding.

PACKAGE THE SKILL
Create SKILL.md in this agent's supported skill location, concise workflow instructions, creative-direction and technical references, portable image-generation (when this agent needs it), video-generation, conversion and validation scripts, and tests for color conversion, metadata, duplicate-submission prevention and secret-safe redirects. Discover dependencies rather than hardcoding machine-specific paths. Run offline tests and a synthetic SDR-to-HDR smoke test without paid generation. Report the installed location, prerequisites and invocation example. If this agent cannot install skills, provide the complete portable skill folder and explain the limitation.

Or download the finished skill

Then ask for a scene

Use the vivid-hdr-video skill to make a 10-second scene of an astronaut walking toward a colossal floating observatory above a turquoise ocean.

What the agent does with that

Composes and inspects the first frame, setting palette, lighting, subject and scale before paying to animate it.
Submits one video job. It checks current models and prices, saves the job ID, and polls until done. If a submission is uncertain it stops instead of paying twice.
Reviews the footage for real footfalls or riding, camera movement, stable anatomy and steady exposure.
Converts the color properly. It decodes SDR to linear light, converts BT.709 primaries to BT.2020, expands selected highlights and encodes HLG. Adding HDR tags alone does none of this.
Packages and verifies. It exports HEVC 10-bit HLG tagged hvc1 plus an SDR comparison, then checks decoding, metadata and raw luma values against a contact sheet and highlight map.

How much does a video cost?

About $1.15 in November 2026: $1.12 for a silent 10-second clip from Kling v3.0 Pro through OpenRouter, plus about $0.03 for the first frame from gpt-image-2.5-sunburst. Prices change, so the skill checks them before it spends anything.

Why does the HDR version look the same, or fail to play?

Your browser or screen may not support HEVC with HLG, or may map it down to SDR. If a file won't play, this page falls back to SDR. Uploading to a social platform can also change the file. We've verified the encoded format here, not how bright it looks on every phone or on X.

What did you change from the original instructions?

The approach is the same: real conversion to HLG, boosted highlights, verification. We didn't crush the shadows, because that made early results look dim. We also treat 75% HLG as reference white under a specific display setup, not as the brightness of an app's interface, and treat highlight percentages as a creative guide, since encoded luma can't tell you what a screen will actually show.

Can I make a GIF instead?

Yes, as a preview. GIF can't carry HDR, so keep the HLG MP4 as the real deliverable.

Ocean Ring

Three steps, in this order

Start with a bright world

Make it really move

Give the highlights room

What you'll need

One prompt to build the skill

Then ask for a scene

What the agent does with that