Communitygithub.com

video-color

Normalize and correct color for a basic edit using verified source signaling, technical transforms, exposure and white balance. Use for SDR/HDR/log handling, neutral shot matching, avoiding double transforms and export color QA.

What is video-color?

video-color is a Claude Code agent skill that normalize and correct color for a basic edit using verified source signaling, technical transforms, exposure and white balance. Use for SDR/HDR/log handling, neutral shot matching, avoiding double transforms and export color QA.

Works with✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/georgewangyu/GeorgeVideoAgent/tree/HEAD/skills/video-color

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Color normalization and correction

Read the project contract. Identify technical source color during intake; creative finishing follows story.

Identify source state [bl: invariant]

Probe each used source for primaries, transfer, matrix, range, bit depth and rotation; inspect a representative decoded sample. Camera name or flat appearance alone does not establish log, HLG or PQ. Metadata can be missing/wrong: record capture declarations and uncertainty. Track originals and converted derivatives separately with transform identities so conversion occurs once.

Default to neutral correction, not a film look or face retouching. If source color is already compatible with the delivery, preserve its technical branch. For HDR/log→SDR, use a verified display transform suited to the actual input, not a metadata-only Rec.709 relabel. Do not guess a camera LUT or stack a LUT on an already normalized copy. Unknown required input color is a specific blocker for faithful normalization; preserve an explicitly provisional ungraded review rather than claim correct SDR.

Correct with bounded samples [bl: invariant]

Name one source→working→delivery branch per used source, with filter/LUT hash, input/output signaling and transform order. Apply technical normalization before small exposure/white balance/contrast adjustments. Preserve highlight detail, blacks and plausible skin; match neighboring shots without forcing different lighting to be identical. A future style pack is optional and cannot replace correct color management. Project instructions can authorize a creative look; base defaults need no individual skill approval.

Use the available editor's color management or installed FFmpeg colorspace, zscale/tonemap or lut3d only with verified support and correct input metadata. Range conversion, transfer conversion and primaries conversion are distinct. Inspect scopes and bounded before/after frames; test every distinct technical branch before a full render. Record identity/pass-through for compatible SDR. Avoid repeated lossy intermediate generations; composite from primary or lossless/mezzanine sources into the final encode.

Verify [bl: invariant]

Check representative bright/dark/skin scenes, adjacent-shot consistency, gradient banding and clipped channels in the exact output. Verify primaries, transfer, matrix and range signaling matches the rendered pixels and intended delivery. Compare actual playback in a color-managed viewer when available; metadata/decode checks alone cannot establish perceptual color fidelity. Return branch/parameter map, sample evidence, final identity and uncertainties to export QA. A new transform invalidates affected visual review, not the story timing.

References

Individual skills in this repo

This repo contains 4 individual skills — each has its own dedicated page.

Related Skills