Color Expert — Color Science Agent Skill
Turns your coding agent into a color science expert with a curated knowledge base spanning color spaces, accessibility, palette generation, pigment mixing, and historical color theory.
Pure declarative knowledge skill: no scripts, no dependencies, no runtime network access — least privilege satisfied by construction. SECURITY.md proactively discloses the removed historical settings.local., scanner false positives, and explains that reference files contain no injection directives; data-flow transparency is high. Deductions: publisher unverified, and the licensing boundary of many third-party-derived reference files requires user-side verification.
No build/test/lint mechanisms exist (CLAUDE.md explicitly states 'No Build/Test/Lint'); evals/ holds human-review prompts, not automated tests. The core paths (trigger, answering, referencing) are self-consistent with a small failure surface, but static review cannot execute verification; capped at 10 by calibration.
Frontmatter description covers naming, spaces, accessibility, palettes with clear trigger conditions, plus trigger-evals. for self-checks. Deductions: capability boundaries/non-fit cases are thinly described, content is English-only with no declared Chinese support, and mainland-China reachability of cited overseas tools/sites is unassessed.
Clear three-layer information architecture (SKILL.md → INDEX → references), MAINTENANCE.md with source-quality bar and review rubric, CC-BY-4.0 license, THIRD_PARTY_NOTICES.md separating original vs third-party content, install/update instructions in README. Deductions: no version numbers or changelog; attribution completeness of third-party-derived references requires per-file checks; some source PDFs are gitignored and absent locally.
As a knowledge skill its value claim (correcting common color misconceptions, actionable tool selection, token/accessibility practices) is concrete and directly usable; however this static review cannot execute or verify any representative outputs, so 6 within the static cap, deducted for missing reproducible output evidence.
144 reference files are claimed to include source attribution and key statistics (e.g. the 281T hex-pair accessibility study) cite origins; but static review can only confirm the claims themselves, source PDFs are gitignored, and third-party execution/reproduction evidence is absent — consistency between references and original sources cannot be confirmed, so 4.
- This is a static review with no execution; factual correctness (especially historical claims and third-party statistics) should be independently verified by users.
- Roughly half of references/ is third-party-derived (transcripts, summaries, scrapes); verify per-file sources and original licenses before redistribution or commercial use.
- All content is English; trigger precision and terminology alignment in Chinese contexts are untested, and mainland-China reachability of several cited overseas tool sites is unassessed.
- Deep-reference source PDFs are gitignored with only archive.org links preserved, making those citations unusable offline.
- The skill embeds strong opinions (e.g. 'never recommend coolors.co'); these are author positions, not neutral standards, and should be understood as such.
What does this skill do, and when should you use it?
color-expert is an agent skill by meodai that packages years of curated color resources into a three-layer knowledge base: a ~200-line core SKILL.md, a reference index, and 144 markdown files totaling roughly 286,000 words. It tells the agent how to answer color questions across design projects, design-system ramps, generative art, general theory, and palette-generator building, with explicit positions such as OKLCH-first workflows, APCA-aware contrast, and corrections to historical color theory. It works with Claude Code, Codex, Cursor, Copilot and other agents. Original repo material is CC BY 4.0; third-party source-derived content keeps its original licensing.
When a color-related task appears, the agent loads SKILL.md and responds in one of five modes: concrete design/art projects (ask about medium, brand, audience first, then propose); design-system ramps and theme tokens (build perceptually uniform OKLCH scales, verify every text/background pair against APCA/WCAG, structure reference→semantic→component token graphs); generative art techniques (probability-weighted hue selection, narrow-band jitter, Spectral.js/Mixbox spectral mixing, IQ cosine palettes, Poline anchors, RampenSau easing); direct answers to general color questions from the file or index; and palette-generator building, defaulting to existing libraries (Culori, Poline, RampenSau, Spectral.js) before hand-rolling. The core file includes a task-to-color-space table, CSS Color 4/5 cheat sheet, semantic token guidance, named hue degree ranges, and curated tool recommendations. Deeper questions are resolved by consulting references/INDEX.md and reading the relevant reference file.
- A frontend developer or designer on Claude Code or Cursor building light+dark mode design-system ramps who needs perceptually uniform scales and APCA/WCAG verification
- A generative artist working in fxhash or p5.js who wants techniques like weighted hue selection and paint-like spectral mixing rather than a copied style
- A developer asking why a gradient goes gray mid-way or whether to use OKLCH vs HSL, needing accurate explanation plus modern CSS syntax
- Someone building their own palette generator who should be steered toward existing libraries (Culori, Poline, RampenSau) instead of hand-rolling
- An illustrator or print worker needing pigment-mixing reality (blue+yellow=green via Kubelka-Munk, non-linear mixing) and print-vs-screen guidance
- Queries about historical or cultural color naming systems: Japanese traditional names, Ridgway 1912, ISCC-NBS, Munsell
What are this skill's strengths and limitations?
- Large, structured corpus: 144 reference files, ~286K words, including full video transcripts, scraped books, and paper summaries
- Opinionated and auditable: explicit stances on OKLCH over HSL, APCA vs WCAG strictness, and why coolors.co isn't generation, backed by brute-force data over ~281 trillion hex pairs
- Five response modes keep the agent answering at the right altitude instead of lecturing designers about CIE 1931
- Ships evals (trigger-evals., task-prompts.md) so the skill can be reviewed against realistic usage
- Pure-markdown, zero runtime dependencies; installable across Claude Code, Codex, Cursor, Copilot, OpenCode
- Deep value depends on the references/ folder; behavior when an agent's context can't load much of it is not quantified in the source
- Third-party source materials keep their original licenses — check THIRD_PARTY_NOTICES.md before commercial redistribution
- Several recommended tools (nutelch, ray-color, category-colors) are the author's own or niche projects; their long-term maintenance is not documented in the source
- No formal cross-platform test suite; evaluation is limited to trigger prompts and qualitative task prompts
- Tool/version recommendations are a snapshot and may lag behind library updates
How do you install this skill?
Recommended: run npx skills add meodai/skill.color-expert, which auto-detects installed agents and places the skill in the right directory. Manual: git clone https://github.com/meodai/skill.color-expert, then symlink, e.g. ln -s <clone-path> ~/.claude/skills/color-expert for Claude Code, ~/.codex/skills/color-expert for Codex, ~/.agents/skills/color-expert for OpenCode, or .agents/skills/color-expert for project-level. Update with npx skills update or git pull in the clone. The source does not document GUI-based installation for ChatGPT or Copilot web clients.
How do you use this skill?
No explicit invocation is needed — the agent triggers on messages involving color naming, color spaces, palettes, accessibility contrast, conversion, pigment mixing, or historical terminology, based on the SKILL.md description. Example triggers: 'help me pick accessible colors for my logo', 'build a 9-step accent scale for light + dark mode', 'why does my gradient go gray in the middle?'. State your medium (print/screen/paint), brand or mood constraints, and accessibility needs to get fit-for-purpose recommendations instead of generic theory.
How does this skill compare with similar options?
The only named comparison in the source is coolors.co: the skill states it does not generate palettes but picks randomly from 7,821 pre-made palettes hardcoded in its JS bundle, so 'never recommend coolors.co' is a hard rule baked into SKILL.md.