Zack Design has published color-grade-ai — a small, practical Claude Code skill that takes a single video frame and generates a targeted .cube 3D LUT to fix the colour problems it sees. It works in DaVinci Resolve, Adobe Premiere Pro, and anywhere else .cube LUTs are honoured.

The pitch, in one paragraph

Feed it a frame. It tells you what is wrong (too yellow, crushed blacks, over-saturated skin, weird magenta cast in the highlights) and generates a .cube LUT that corrects it. Drop the LUT onto a node in Resolve, a Lumetri Look slot in Premiere, or an ffmpeg -vf lut3d pass — and the fix is applied.

How it works

Under the hood, color-grade-ai is a pair of small Ruby + Python scripts with a LUT-chain engine sitting between them:

# Auto-analyse a frame and get per-node recommendations
python3 auto_grade.py frame_709.png

# Generate a single correction LUT from a named preset
ruby generate_lut.rb red_skin_fix skin_fix.cube

# Bake a creative chain into one LUT
ruby generate_chain_lut.rb studio_balanced.cube \
  [email protected] [email protected] [email protected] [email protected]

The @0.8 syntax is how each node in the chain carries its own intensity, so you can dial in a look without editing the underlying preset. When a chain is baked into a single .cube, it loads as one LUT in any editor — no node-tree rebuilding required.

Interactive preview

Because nobody wants to bounce between an editor and a terminal, the repo ships a browser preview UI. Drag a frame in, load a conversion LUT, dial corrections across a six-node chain, and see the result live. It is a ~150 line HTML page backed by a local Python http.server — no build step, no framework.

As a Claude Code skill

color-grade-ai is structured as a Claude Code skill, which means once it is cloned into ~/.claude/skills/color-grade (or the project-level equivalent), you can ask Claude things like “grade this frame for a warm studio look, keep the skin neutral” and it drives the same scripts with the right presets, chains, and intensities. The SKILL.md file documents every preset, every node in the chain, and every creative preset so the model has a structured API surface to reason about.

Why it is useful

Most editors grade by feel — curves, wheels, a faint sense that “this is too yellow.” That is a perfectly fine workflow for a hundred clips. It is not a fine workflow for a couple of hundred clips shot across multiple cameras under mixed light, where consistency matters more than feel. color-grade-ai exists for the batch case: give it the same analyser and the same LUT-chain, and every clip in a folder gets graded against the same reference.

Source on GitHub, full docs at isaacrowntree.github.io/color-grade-ai.