🖍️ color-correct

An AI colorist skill for Claude Code — your agent looks at your footage, diagnoses it, grades it, and shows you a before/after before rendering. Nine named looks reverse-engineered from creators with great color. ffmpeg-based, non-destructive.

What it does

• Pick a look — "apply Oat Milk to this clip"
• See the menu — "show me all the looks on this video" → labeled 3×3 contact sheet on your own footage
• Let AI decide — "make this look good" → it diagnoses (talking head vs b-roll vs screen recording), fixes casts/exposure first, picks a fitting look, and shows a side-by-side before committing
• Steal any look — "make my footage look like [creator]" → consensus fingerprinting across multiple videos separates the grade from the scenery

The nine looks

Every look ships at subtle strength by default — graded footage should look shot well, not filtered.

🧊 LUTs included — use the looks anywhere

Every look also ships as a standard .cube 3D LUT (36³), so you can use them without an agent — in Premiere, DaVinci Resolve, Final Cut, CapCut, or straight ffmpeg:

luts/subtle/   ← the shipping strength (start here)
luts/full/     ← full consensus strength (flat footage / "more")

ffmpeg -i in.mp4 -vf "lut3d=luts/subtle/golden-hour.cube" -c:a copy out.mp4

LUTs are generated from the exact same chains via scripts/gen_luts.sh (HALD-CLUT method, verified to match within ~1/255).

How the looks were made

Not vibes — measurement. For each reference creator we sampled ~18 frames across 3 different videos (different scenes, lighting, even climates), computed a per-frame grade fingerprint (luma percentiles, per-tonal-band color casts, saturation), and took the median + IQR. Tight IQR across all scenes = the actual grade; wide IQR = scenery, discarded.

The single-frame version of this gets it wrong in fascinating ways — one creator's "cool daylight look" turned out to be a blue sky, and their actual grade was warm. The consensus method (and that story) is documented in SKILL.md.

Looks studied from creators we admire: Elias & Kajsa, Benji Plant, Jonna Jinton, Meek Surf, Radhi Devlukia, Bryan Johnson, Enrico Tartarotti, David Dobrik, Paige Wassel. 💛 These are interpretations of their color character, derived from published videos — go watch them.

Install

git clone https://github.com/louisedesadeleer/color-correct.git

Then point your agent at it. In Claude Code, add to your CLAUDE.md:

#### /color-correct
Read: /path/to/color-correct/SKILL.md
Grade a video with Claude as the colorist — pick a named look, see the menu, or let AI decide.

Requirements: ffmpeg, python3 + Pillow/numpy (pip install pillow numpy), yt-dlp (only for Steal mode).

Usage

/color-correct path/to/video.mp4                     → AI decides
/color-correct path/to/video.mp4 apply Golden Hour   → one look
/color-correct path/to/video.mp4 menu                → 3×3 sheet of all looks on your frame
/color-correct video.mp4 --like <youtube-url>        → steal a new look

LUTs + agent: who does what

The LUTs are the deliverable. The agent is the colorist. They're layers, not rivals:

• A LUT is fixed math — maximum compatibility, same result everywhere. Perfect for distributing a look into any tool.
• But a LUT can't correct: apply Golden Hour to footage with a green webcam cast and you get a beautifully graded green cast. White balance and exposure are footage-dependent by definition — no file can contain that decision.
• The agent does everything a file can't: looks at your footage, fixes casts and exposure first, picks a look that fits the content, adapts the strength when a look fights your lighting, skips the screen-recording b-roll, and re-measures after grading to verify the numbers moved where predicted.

Use the LUTs alone and you get the looks. Use the agent and you get a colorist who happens to carry LUTs in their pocket.

Built by Louise de Sadeleer with Claude. Part of a series of AI video-editing skills.