color-expert: a 200-line entry point in front of 286,000 words

A 200-line entry point in front of 286,000 words

The design is a progressive-disclosure stack with a very steep ratio. `SKILL.md`, about 200 lines, is the layer the agent loads immediately and is described as the greatest hits: key facts, corrections, tool recommendations and the guidelines that answer most colour questions without digging further. `references/INDEX.md`, about 220 lines, is a structured lookup table the agent scans to pick a file. Behind that sit 144 Markdown files and roughly 286,000 words, described as full video transcripts, article summaries, library documentation, scraped websites and research notes. So the always-loaded surface is on the order of four hundred lines and everything else has to be found. That is the right shape for a context budget, and it puts the whole burden on INDEX.md being accurate, since a reference that is not in the table is a reference the agent will not open.

The evals folder holds prompts, not evaluations

There are two eval files and no runner. `evals/trigger-evals.json` is for sanity-checking whether the frontmatter description is likely to trigger in the right situations, and `evals/task-prompts.md` is for testing whether the skill answers realistic colour questions at the right level. Both are inputs to a judgement a human makes, and neither is executed, scored or given a pass condition. The README is upfront about this, describing the setup as a minimal review loop rather than a heavy benchmarking framework, which is honest framing for a knowledge base whose main risk is unverified recall. Two companion documents fill the rest of the process: `MAINTENANCE.md` decides whether something belongs in `SKILL.md`, in `references/`, or not in the repository at all, and `ROADMAP.md` tracks planned colour-specific scripts. Neither the trigger file nor the task file has a recorded outcome anywhere in the visible documentation.

CC-BY-4.0 over a corpus assembled by scraping two named sites

The licence choice fits the artifact: a knowledge base of prose is content, and CC-BY-4.0 is a content licence. The repository also carries `THIRD_PARTY_NOTICES.md` and `SECURITY.md` beside the LICENSE, which is the right hygiene for something assembled this way. What the capture workflow actually does is the thing to understand. Videos are transcribed with yt-dlp, summarised and mined for concepts. PDFs and documents are converted to Markdown with markitdown. GitHub repositories have their READMEs and docs fetched. Articles have their content extracted. Books mentioned in videos are searched on Archive.org and freely available PDFs are downloaded, with fourteen of those kept locally and gitignored. Websites are fully scraped chapter by chapter, and two are named: huevaluechroma.com, eleven chapters plus a glossary, and colorandcontrast.com, whose content is described as extracted from an SPA bundle. An open licence on the compilation is not an open licence on transcripts of other people's work or on a chapter-by-chapter copy of a site.

The accessibility headline number is somebody else's brute-force run

The most quotable figure in the skill is a scan rather than a specification. Under accessible colour pair research, the entry credits a Rust brute-force run over roughly 281 trillion hex pairs and reports that only 11.98 percent pass WCAG AA while 0.08 percent pass APCA 90. That is a useful calibration of how tight those thresholds are, and it is also someone else's computation, carried into the skill as a fact without a method description, a date or a link to the run. The skill's own accessibility tooling is a separate list: APCA and Myndex, described here as the WCAG 3 algorithm, apcach for contrast-first colour composition, and Bridge-PCA. Those are the entries an agent would be pointed at; the 281 trillion figure is context. Worth keeping the two apart when deciding which to trust when an answer turns on it.

Three named sections account for 119 of the 144 files

The numbers table gives 144 Markdown reference files and roughly 286,000 words, plus 14 local gitignored source PDFs, 48 catalogued online tools and 54 or more transcribed video sources. The three sections named in the visible text are historical colour science at 14 files, contemporary colour science at 55 and techniques, tools and libraries at 50. That totals 119, leaving 25 files unattributed, and the table breaks off mid-sentence in the practical-design entry, so the remaining sections are presumably described further down. The structure block makes the same point from the other direction:

SKILL.md                              # The skill definition (loaded on activation)
CLAUDE.md                             # Claude Code repo instructions
references/
  INDEX.md                            # Master lookup table
  historical/                         # Pre-digital color science
    *.md                              # Ostwald, Helmholtz, Bezold, Ridgway, ISCC-NBS,
                                      # Moses Harris, Amy Sawyer, Lewis/Ladd-Franklin,
                                      # Caravaggio's pigments, Itten critique...
    pdfs/                             # Sourc

It shows `SKILL.md`, `CLAUDE.md`, a `references/` directory with `INDEX.md` and `historical/`, a glob naming the historical authors, and a `pdfs/` line whose own comment stops mid-word. So the layout is documented for one of the three folders the capture process writes into.

The historical shelf is the part organised as an argument

Fourteen files carry most of the skill's distinctive value, and they are listed with dates and a verdict attached, which is what separates them from a reading list. Moses Harris in 1769 is credited with the origin of bad RYB colour theory, on the grounds that his own wheel needed a fourth pigment. Bezold in 1874 killed indigo as a spectral colour and renamed it ultramarine blue; Helmholtz in 1856 is the last major physicist to use the word. Amy Sawyer patented a CMY wheel in 1911, decades before it was mainstream, and Elizabeth Lewis married trichromatic and opponent-process theory in 1931, anticipating CIE Lab by thirty years. Ridgway's 1912 system of 1,115 named colours is noted as fully digitised as JSON, and ISCC-NBS from 1955 as 319 systematically named blocks. Ostwald's 24-hue circle from 1918 to 1930 dominated the 1930s and 40s and then disappeared.

The tool entries carry other people's download counts

Fifty files sit on the practical side, and the attribution is thorough: named tools for palette generation including RampenSau, Poline, pro-color-harmonies, dittoTones, FarbVelo, an IQ cosine formula and CSS-native generation with `color-mix()`. Libraries come with their characteristics, Culori for 30 spaces and 10 distance metrics, Color.js as the CSS spec editor, Spectral.js for Kubelka-Munk, RYBitten for 26 historical cubes. Analysis and linting entries carry rule counts, 38 for Color Buddy and over 20 visualisation widgets for Censor. Some of those numbers are the skill's own measurements and some are borrowed claims: Color.js is described with 154 million downloads, @texel/color as 5 to 125 times faster, and colornerd as holding 29,875 manufacturer swatches. Naming the source is the right practice; the counts still age at the speed of the upstream projects.

Editorial conclusion

Use color-expert when your agent needs colour science that is current and cited rather than recalled, because the historical shelf alone is organised as a chronology with arguments attached, which is the kind of thing a model will not produce from memory. Two things to weigh first. The corpus was assembled by transcribing videos and scraping two named websites chapter by chapter, so the CC-BY-4.0 grant covers the author's own compilation and a `THIRD_PARTY_NOTICES.md` sits alongside it, but the upstream rights in the transcripts and scrapes are not something the repository can grant. And the evals folder is prompt fixtures, not a test suite, so there is no measurement of whether the skill triggers when it should or answers at the right depth.

Frequently asked questions

Official sources

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