taste-skill

What It Does

taste-skill is a community skill pack by Leonxlnx that makes coding agents produce frontends that do not look like default LLM output. It ships 13 Agent Skills format SKILL.md prompt contracts in two groups: 10 implementation skills (the v2 default design-taste-frontend, its preserved v1 pin, a stricter GPT/Codex variant, an image-to-code pipeline, a redesign auditor, plus style-specific contracts for soft, minimalist, brutalist, and Stitch workflows, and an output-length enforcer) and 3 image-generation skills that produce design reference boards for web, mobile, and brand kits. The user installs one or more skills into their coding agent (Codex, Cursor, Claude Code, or any agent supported by the skills CLI), and the loaded contract steers layout, typography, motion, and spacing decisions at generation time. It targets developers who keep getting generic AI-purple gradient landing pages and want the agent to read the brief first.

The Good

The v2 core skill is the best-written prompt contract in this tier GearScope has reviewed. The default skill (design-taste-frontend, 1,206 lines, 18 numbered sections plus 3 appendices) opens with a Brief Inference protocol: before any code, the agent must declare a one-line design read ("Reading this as: B2B SaaS landing for technical buyers...") and ask at most one clarifying question. Section 1 defines three 1-10 dials (DESIGN_VARIANCE, MOTION_INTENSITY, VISUAL_DENSITY) with two inference tables mapping vibe words and use cases to dial values, so configuration is conversational instead of file editing. Section 2 is a design-system honesty map: 12 brief patterns route to official packages (@fluentui, @carbon, govuk-frontend, uswds, shadcn/ui) with an explicit rule never to hand-copy a system's CSS, while 8 aesthetics (glassmorphism, bento, brutalism) are labeled as having no official package, and Apple Liquid Glass is documented as a web approximation rather than pretended to be a library.

The AI Tells catalog reads like it was written from production scar tissue, because it was. Section 9 bans dozens of concrete failure signatures with specifics instead of adjectives: version labels in heroes (V0.6, INVITE-ONLY PREVIEW), section-number eyebrows (001 · Capabilities), div-based fake product screenshots ("the #1 LLM-design Tell"), "Quietly in use at" social-proof headers, locale and weather strips (LIS 14:23 · 18°C), decorative status dots, photo-credit captions as decoration, and startup-slop brand names (Acme, Nexus, SmartFlow). The em-dash ban (Section 9.G) is phrased as binary and non-negotiable, with an honest rationale in the text: "The agent has historically ignored em-dash limits when phrased as 'use sparingly.'" Section 14 closes with a 50-item mechanical pre-flight checklist where every box must pass before shipping, including countable rules (eyebrow count at most ceil(sectionCount / 3), no two CTAs with the same intent). Section 13 declares what the skill is not for (dashboards, data tables, multi-step forms), which most competing packs never state.

Every install path lands byte-identical files, verified four ways. We ran the documented skills CLI install in isolated scratch directories: single-skill (design-taste-frontend), the v1 pin (design-taste-frontend-v1), the full pack, and the one install name whose folder differs most from its install name (image-to-code). All four exited 0, all landed at the hermes-agent spot .hermes/skills/ with a skills-lock.json, and every landed SKILL.md was byte-identical to repo HEAD (no stale registry blob, the failure mode we caught on hyperframes). The full-pack install landed exactly 13 of 13 skills. The Claude Code plugin manifest passed claude plugin validate --strict clean.

Demand is exceptional and broad: all 13 skills chart on skills.sh, totaling 3,464,236 installs. Per-skill all-time counts pulled live from the skills.sh API: design-taste-frontend 404,917; high-end-visual-design 301,638; redesign-existing-projects 297,954; minimalist-ui 276,251; full-output-enforcement 263,079; industrial-brutalist-ui 257,028; stitch-design-taste 255,585; gpt-taste 254,913; brandkit 244,910; image-to-code 239,322; imagegen-frontend-web 238,832; imagegen-frontend-mobile 233,805; design-taste-frontend-v1 192,329. No skill in the pack is dead weight; even the 13th-ranked skill has 192K installs. The 80,799 GitHub stars (API-verified) and 5,538 forks sit on 154 commits from 6 contributors going back to 2026-02-19.

The version transition is handled the right way. The v2 rewrite ships as the default while v1 stays installable under a separate pin name (design-taste-frontend-v1), the README documents both, and the CHANGELOG explains the split and the rationale with a per-section diff of what v2 adds. The output-skill is backed by a real research tree (research/laziness/ with root-causes, remediation, and findings subdirectories citing studies) rather than vibes.

The Bad

The repo's own surfaces disagree about skill names. The README says the install name (the frontmatter name field, e.g. image-to-code) is "the exact value you pass to --skill", and the skills CLI honors that. But the skill.sh helper script registers 9 of 13 skills under folder names instead ([image-to-code-skill], [redesign-skill], [soft-skill], [output-skill], [minimalist-skill], [brutalist-skill], [stitch-skill], [taste-skill], [taste-skill-v1]), so sourcing it with the documented install name fails for most of the pack. skills/llms.txt has the inverse wrinkle: 12 entries keyed by folder name, one (gpt-taste) by install name. The skills.sh registry also still serves a legacy entry leonxlnx/taste-skill/image-taste-frontend (3,673 installs) for a skill that no longer exists in the repo. None of this breaks the documented install path, but a pack whose whole premise is precision ships three inconsistent naming schemes.

The default install is an experimental pre-release, and there is no release to pin to. The CHANGELOG lists v2 under [Unreleased], the skill text itself says "Actively iterating toward v2.0.0 stable", and the repo has zero git tags and zero GitHub releases despite 154 commits and 80K stars. The plugin manifest has been pinned at 1.0.0 throughout. Anyone who wants deterministic skill content across a team has exactly one stable anchor: the v1 pin. Everything else floats with main.

No CI, no tests, no eval harness. The naming drift above is exactly the class of defect a 20-line structural check in CI would catch (frontmatter name vs README table vs skill.sh vs llms.txt), and it shipped anyway. For a prompt-contract pack the absence of runtime tests is normal; the absence of even structural linting at this scale (3.4M installs) is a choice worth noting, and it contrasts with peer packs that run install-name consistency checks in CI.

Hermes and OpenClaw users are undocumented despite a fully working install. The README names Codex, Cursor, Claude Code, and ChatGPT; grep finds zero real Hermes or OpenClaw mentions in the README or any skill (the single hit is the Hermès luxury brand inside a color-palette description). Our hermes-agent install worked perfectly, so the gap is documentation, not compatibility. Related: the repo ships a valid .claude-plugin/marketplace.json that passes strict validation, but the README never mentions the Claude Code plugin-marketplace path at all; npx skills add is the only documented install.

The README front-loads monetization above all content. Before a single word about skills, the reader gets a full-width Kimi (Moonshot AI) affiliate banner with tracking IDs and a "10% bonus API credits" promo, then a 7-row sponsor table. The last six commits at review time were all sponsor-layout work, not skill iteration. None of this affects the skill files, and the license is clean MIT, but readers should know the README leads with ads. The README also carries a "no official token, coin, or crypto project" disclaimer, which tells you the repo's popularity has attracted impersonators; treat any taste-skill token as a scam.

Smoke Test Results

Testing ran on the host (macOS 26.5.2, aarch64, Node 24.13.1, npx from /usr/local/bin). The repo was cloned with git clone --depth 1 (3.3 MB, 13 skill directories). GitHub stats were API-verified (80,799 stars, 5,538 forks, created 2026-02-19, pushed 2026-08-24, MIT, 58 open issues, 154 commits, 6 contributors, no tags). skills.sh install counts were pulled live from the registry API.

Run A. Fresh clone, structural validation

$ bash structure.sh # 20 checks over the fresh clone (frontmatter, names, manifests, links)
PASS - 13 SKILL.md files present, none duplicated by platform trees
PASS - all 13 have name + description frontmatter
PASS - 13 frontmatter install names match the README install-name table exactly
PASS - skill.sh registry paths all resolve (13 entries)
NOTE - skill.sh keys 9 of 13 by folder name where README documents install names
PASS - plugin.json + marketplace.json parse as JSON
PASS - LICENSE present (MIT)
PASS - skills/llms.txt indexes all 13 skills (mixed folder/install keying noted)
PASS - all descriptions under the 1024-char limit
PASS - v2 SKILL.md carries 18 sections + appendices
PASS - all three dials defined (DESIGN_VARIANCE / MOTION_INTENSITY / VISUAL_DENSITY)
PASS - Section 9.G em-dash ban present
PASS - Section 14 pre-flight check present
PASS - stitch-skill DESIGN.md companion present
PASS - research/laziness tree present (backs output-skill)
PASS - README relative file links resolve
PASS - taste-skill-v1 documents its pin purpose
PASS - CHANGELOG documents the v2 experimental split
PASS - no API keys or tokens committed in skills/, scripts/, skill.sh
PASS - README example images exist

Pass rate: 20 of 20. Structure is clean; the naming inconsistencies are recorded as NOTES rather than failures because the documented install path (the skills CLI) uses the correct frontmatter names throughout.

Full structural log ->

Run B. Install verification (the documented functional surface)

Run in isolated scratch directories with an isolated HOME so nothing landed in the real ~/.hermes. Scripted flags per the README's own non-interactive guidance.

$ npx -y skills add https://github.com/Leonxlnx/taste-skill --skill design-taste-frontend --agent hermes-agent --copy --yes
PASS - exit 0; landed at .hermes/skills/design-taste-frontend/SKILL.md

$ diff repo/skills/taste-skill/SKILL.md .hermes/skills/design-taste-frontend/SKILL.md
PASS - byte-identical to repo HEAD (no stale registry blob)

$ ls .hermes/skills/skills-lock.json
PASS - lockfile written for reproducible installs

$ npx -y skills add https://github.com/Leonxlnx/taste-skill --skill design-taste-frontend-v1 --agent hermes-agent --copy --yes
PASS - v1 pin install works; landed file byte-identical to repo v1

$ npx -y skills add https://github.com/Leonxlnx/taste-skill --agent hermes-agent --copy --yes
PASS - full-pack install lands 13 of 13 skills

$ npx -y skills add https://github.com/Leonxlnx/taste-skill --skill image-to-code --agent hermes-agent --copy --yes
PASS - install name (not folder name) resolves, landing image-to-code/SKILL.md

$ claude plugin validate --strict .
PASS - validation passed on .claude-plugin/marketplace.json

Pass rate: 7 of 7. The documented install surface is solid in every variant tested, including the v1 pin and the folder-name-differs case.

Full functional log ->

What the runs tell you

The pack is structurally clean and the install path is the strongest part of the engineering: four install variants, all byte-identical to the repo, with a lockfile. What no host run can verify is the thing users actually install it for: whether pages generated under the contract look better. That requires live agent sessions against real briefs, which this cycle did not run; the 3.4M-install demand and the specificity of the rule text are the available evidence for the output claim.

Setup Walkthrough

1. Pick the skill you want from the README table; the safe default is design-taste-frontend (the v2 experimental rewrite).
2. Run npx skills add https://github.com/Leonxlnx/taste-skill from your project directory and pick skills at the prompt, or script it: npx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend" --agent --copy --yes.
3. If you need the pre-v2 behavior pinned, install design-taste-frontend-v1 the same way; both can coexist.
4. Zero dependencies, zero API keys, no build step. The skills are markdown contracts; the agent loads them on demand.
5. Gotchas: the three dials are overridden conversationally (the skill forbids editing the file), image-generation skills produce reference images only (no code), and Section 13 of the v2 skill puts dashboards, data tables, and multi-step forms out of scope.

Alternatives

1. pbakaus/impeccable (4.5, reviewed 2026-08-04) - a 23-command design system with ~59 deterministic, no-LLM anti-pattern detectors and browser iteration. Choose impeccable when you want to lint and fix existing HTML; choose taste-skill when you want to prevent the slop at prompt time.
2. anthropic's frontend-design skill - the origin of the tier and the shorter, less opinionated contract. taste-skill is longer, more current on LLM-specific tells, and dial-configurable; the Anthropic skill is the conservative baseline.
3. google-labs-code/stitch-skills (4, reviewed 2026-05-18) - the taste-design skill there is Stitch-bound; taste-skill's stitch-design-taste covers the Stitch case inside a pack that also serves non-Stitch workflows.
4. greensock/gsap-skills (4.5, reviewed 2026-08-03) - orthogonal and complementary: GSAP correctness for animation code specifically, where taste-skill governs overall design direction.