meodai/skill.color-expert: an agent skill that ships 286,000 words of color science

What skill.color-expert solves, and who it is for

Coding agents answer color questions with confident approximations. Ask one why a gradient looks muddy in the middle and you get a plausible paragraph about RGB interpolation that may or may not mention perceptual spaces, may cite a color model that was superseded, and will rarely name a source. skill.color-expert exists to replace that with curated material. The README describes it as "an agent skill that turns your coding agent into a color science expert," built from resources the author kept looking up and sharing.

The audience is narrow and specific. It is for people using Claude Code, Codex, Cursor, Copilot or another agent that supports the agent skills format, and who work on color as a craft problem: picking palettes, checking contrast, mixing pigments, explaining why a historical color wheel was wrong. It is not for someone who wants a color conversion library to import. Nothing in the repository is a runtime dependency. The deliverable is text that the agent reads before it answers.

The three-layer structure and how an agent loads it

The design is a retrieval hierarchy, and the whole value proposition depends on the agent respecting it. Layer one is SKILL.md, roughly 200 lines, described in the README as the "greatest hits" that the agent loads immediately: key facts, corrections, tool recommendations and guidelines that answer most questions without going deeper. Layer two is references/INDEX.md, about 220 lines, a structured lookup table the agent scans to find the right reference file for a topic. Layer three is the references/ directory itself: 144 markdown files, roughly 286,000 words, holding video transcripts, article summaries, library documentation, scraped websites and research notes.

That ordering matters because the cheap layer is the one that always loads. If a question can be answered from SKILL.md, no reference file is read and no context is spent. The index is the routing decision. The deep files are pay-per-use. An agent that ignores the index and reads reference files at random will burn context on 286,000 words of material it does not need, which is the failure mode this structure is meant to prevent. The repository also carries an evals/ folder with trigger-evals.json for checking whether the frontmatter description fires in the right situations, and task-prompts.md for qualitative review of answers.

Installing skill.color-expert and asking a first color question

The README does not print an install command. It points to agentskills.io as the agent skill format and describes the repository contents, so installation follows whatever mechanism your agent uses for skills. The practical route is to place the repository where your agent scans for skills and let it pick up SKILL.md.

After that, the top level should contain SKILL.md, CLAUDE.md, MAINTENANCE.md, ROADMAP.md, evals/ and references/. The repository layout in the README shows the evals/ folder holding two files:

evals/
  trigger-evals.json                  # Realistic should-trigger / should-not-trigger prompts
  task-prompts.md                     # Realistic color tasks for qualitative review

The first file holds should-trigger and should-not-trigger prompts for the frontmatter description. The second holds color tasks for review. A first useful test is a question the skill was built to answer better than a generic model. Ask your agent something like: why does interpolating a gradient in sRGB produce a dark band between two saturated colors, and which space should I use instead. If the skill is active, the answer should draw on the OKLAB material the README lists among the contemporary references. Run your own prompts against both eval files before trusting the skill in a real project.

The pdfs/ subfolders under references/historical and references/contemporary are gitignored and about 236MB in the author's working copy, so a fresh checkout will not have them. The markdown that summarizes those sources is present.

Where the knowledge comes from, and what that implies about accuracy

The reference set is opinionated by construction. The historical folder covers Ostwald's 24-hue perceptual circle, Helmholtz on physiological optics, Bezold's removal of indigo as a spectral color, Ridgway's 1,115 named colors digitized as JSON, the ISCC-NBS 319 named blocks, Moses Harris and the origin of RYB theory, Amy Sawyer's CMY wheel patent, and Elizabeth Lewis's 1931 work that the README says anticipated CIE Lab by 30 years. The contemporary folder holds Bjorn Ottosson's four OKLAB posts, a full scrape of David Briggs's huevaluechroma.com across 11 chapters plus glossary, colorandcontrast.com extracted from its SPA bundle, and 20-plus Color Nerd video transcripts. The techniques folder lists Spectral.js, Culori, Color.js, RampenSau, Poline, RYBitten, PickyPalette, Color Buddy lint rules, APCA/Myndex, the IQ cosine formula, Cubehelix, Tyler Hobbs on generative color, pixel art palettes, Book of Shaders and LYGIA.

That is a strong syllabus, and it is also a single curator's syllabus. The README is explicit that the collection grew from what the author kept pasting and referencing, and that Itten's seven contrasts appear as a "critically reviewed" entry rather than a neutral summary. So the skill carries the author's judgments along with the sources. For a topic like pigment mixing or historical color naming, that is useful, because the disagreements are the interesting part. For anything where you need a defensible standard citation, treat the reference files as pointers to the underlying papers and videos, not as the authority itself.

The limits: no code, no versions, and a large surface to maintain

The clearest limitation is that skill.color-expert cannot compute anything. It has no color conversion function, no contrast checker, no palette generator. If your build pipeline needs to convert OKLCH to sRGB at runtime, this repository will not do it, and the correct move is to read its notes on Culori or Color.js and then depend on that library directly. Using the skill as a substitute for a library is the wrong tool.

The second limit is version drift. The references document tools and libraries as they were when captured. Nothing in the repository pins a version of Culori, Color.js or Spectral.js, and there are no releases to track. If a library changes its API, the reference file describing it does not update on its own.

The third is maintenance cost, which falls on whoever forks it. The README lays out an ingestion workflow: videos transcribed with yt-dlp, PDFs and documents converted with markitdown, GitHub repos documented from their READMEs, articles extracted, books searched on Archive.org, websites scraped chapter by chapter, and tools collected into an Online Tools table. MAINTENANCE.md defines what belongs in SKILL.md versus references/ versus nowhere, and a review rubric. That is a real editorial process, and it is the part most people underestimate. Adding a source is easy. Deciding whether it belongs in the always-loaded layer is not.

How it compares to color libraries and to plain model knowledge

The obvious alternative is a color library such as Culori or Color.js, and the difference is categorical rather than competitive. A library exposes functions with defined inputs and outputs; you call it from code and get a number back. skill.color-expert exposes prose that an agent reads before writing an answer. If you ask a library why a gradient looks muddy, it returns a conversion, not an explanation. If you ask an agent with the skill loaded, you get an explanation that may cite Ottosson, Briggs or the APCA material, and then you still need a library to apply it. The two are complementary, and the repository's own techniques folder documents several of those libraries, which is a reasonable sign that the author sees them the same way.

The other alternative is doing nothing and relying on the model's built-in color knowledge. The gap shows up on specifics. A general model will usually get the broad strokes of OKLAB and WCAG right and will often be vague or wrong about Ridgway's color names, the ISCC-NBS blocks, the status of indigo, or the details of Caravaggio's copper resinate technique. Whether that gap matters depends entirely on the questions you ask. For a contrast ratio check, the model is probably fine. For a historical palette reconstruction or a pigment mixing explanation, the reference set is the reason to install it.

Licence, upgrade path and what to check before you commit

The repository is licensed CC-BY-4.0, which is a content licence rather than a software licence. Attribution is a condition of reuse, and the repository ships a THIRD_PARTY_NOTICES.md file, which is consistent with a collection assembled from other people's videos, papers, articles and websites. If you redistribute the reference material, or build a product on top of it, read the licence text and the notices file rather than assuming a permissive code licence applies. This is a description of what the files say, not legal advice.

There are no releases, so there is no upgrade path in the usual sense. The last push was on 2026-09-02, and the default branch is main. Updating means pulling the branch and re-reading whatever changed. Because SKILL.md is the layer that always loads, a change there affects every answer the skill produces, which makes it the file to diff first. There is also a ROADMAP.md listing planned scripts and future extensions, and a SECURITY.md at the top level. Check the roadmap before building your own tooling around the repository, since planned scripts may overlap with what you intend to write.

Editorial conclusion

Adopt skill.color-expert if your agent already answers color questions and you want those answers grounded in named sources: OKLAB posts, huevaluechroma, APCA and Myndex material, Ridgway's digitized names. Skip it if you need a deterministic palette engine with pinned versions and reproducible output, because this repository holds prose, not code. Before installing, open SKILL.md and confirm the frontmatter description triggers on your prompts, check that references/INDEX.md routes to the topics you care about, and read MAINTENANCE.md for the source quality bar. The license is CC-BY-4.0, so attribution travels with the content.

Frequently asked questions

Official sources

1. Issues
2. License: CC-BY-4.0
3. meodai/skill.color-expert on GitHub
4. README