Taste Skill Framework (Leonxlnx/taste-skill)

Open-source anti-slop framework for AI agents. Source: github.com/Leonxlnx/taste-skill + tasteskill.dev (Leon Brown / Leonxlnx). The upstream collection is preserved as evidence, while Kevin-Wiki exposes one active executable owner: skills/engineering/taste/SKILL.md.

Why It Exists

LLMs default to the statistical center of their training data. Without hard constraints, every AI-generated frontend collapses into the same look: Inter font, centered hero, blue gradient, three equal cards, blur blobs. The taste-skill framework is a shared rule library that makes generic output structurally impossible - variance dials force asymmetry, banned patterns block cliches, required UI states catch missing loading/empty/error coverage. Source: https://github.com/Leonxlnx/taste-skill; https://www.tasteskill.dev/

The framework is shared across Anthropic Claude, GPT/Codex, Cursor, and Google Stitch via skill files plus project rules. Same constraints, same vocabulary, same anti-pattern list.

Current Local Topology

Kevin-Wiki does not install the upstream files as thirteen peer capabilities. The active runtime owner is the consolidated taste skill at skills/engineering/taste/SKILL.md; former upstream routes are folded references loaded only when the task needs a specific visual mode or workflow modifier. This preserves the full source vocabulary without making agents choose among a crowded list of near-duplicate skills.

The upstream repository remains a source collection, not the local registry contract. New upstream material must first be reviewed for a distinct job, then either improve the consolidated owner, remain a cited reference, or be held. Do not bulk-copy the repository into the runtime.

Components

Image Generation Route

The May 2026 bookmark points at skills/imagegen-frontend-web/SKILL.md: use the taste framework before code by generating one image/reference per website section, then implementing against the visual direction. The durable pattern is reference board first, UI code second for high-taste web work. This belongs in the taste-motion/design workflow, not as a generic image tool, because the skill's value is the design constraint system around the images. Source: X/@LexnLin, 2026-05-02; Source: Leonxlnx/taste-skill, 2026-06-30

The reviewed video artifact shows the rule in practice: one "Forma" site concept rendered as multiple horizontal section references - hero, system/section cards, testimonial proof, and closing CTA - instead of one tall stitched mockup. That makes the design reviewable and codeable section by section, which is the exact opposite of asking a model to hallucinate a finished webpage from a single mood image. Source: X/@LexnLin, 2026-05-02; Source: local video contact sheet, reviewed 2026-07-02

Current source snapshot reviewed on 2026-08-10: default-branch HEAD e988add20dab0fa97d7a76781c48961c8184288e, MIT license, 75,029 GitHub stars, and 5,146 forks. The prior Brin result remained caution with medium confidence, so upstream changes remain source snapshots until a fresh audit passes; reputation is evidence of signal, not permission to overwrite the local skill. Source: X/@nurijanian, 2026-07-10; Source: captured GitHub repository and README, 2026-08-10

Brand Kit Route

The DrapzDZN Radiant artifact is a useful taste-brandkit benchmark because it shows a full identity board rather than a lone logo: symbol and wordmark, spacing grid, primary logo, Host Grotesk typography, brand colors, logo usage, poster/social applications, and a coherent dark luminous wave motif with peach accent applications. The review standard for AI-generated identity work is therefore system completeness: capture logo geometry, type, palette, usage, and applications before treating a brand direction as reusable. Source: X/@DrapzDZN, 2026-03-14; Source: local image artifact, reviewed 2026-07-02

The Second Eight Radial artifact extends that benchmark for technical/health AI identity. The local board combines logo construction, biological/scientific textures, data coordinates, typography specimens, UI shells, workflow cards, color chips, and product copy around "orchestration" for health and life sciences. The reusable taste-brandkit rule: a technical brand kit should prove semantics through system diagrams, UI fragments, and domain-specific texture, not just an abstract mark. Source: X/@secondeight_, 2026-04-09; Source: local image artifact, reviewed 2026-07-03

The jamescoder12 Gemini prompt thread is useful as a prompt checklist, not as proof that logo designers are obsolete. Its seven-step sequence covers brand foundation, logo concepts, color psychology, typography, visual direction, voice/messaging, and a complete brand bible. The reusable rule for taste-brandkit: ask for the brand system in layers, then judge the output against real guideline references such as Public Brand Guideline References. A single pretty logo concept is not enough to promote a brand direction. Source: X/@jamescoder12 thread, 2026-03-12

The August Runline/SYNC/Drapz replay corrects one deeper flaw in the upstream instruction: a complete brand world cannot be proved by one overview image. Aron's Runline post says the team spent weeks on research, moodboards, and directions before selecting the identity; Pixel Orb's live studio process then separates exploration, refinement into a scalable system, and launch delivery. The studio and Runline product surfaces identify the actual deliverables as branding, web, graphic design, product-mechanism diagrams, risk/compliance surfaces, content, and launch collateral. taste-brandkit now requires an evidence brief, at least two materially distinct directions, an explicit selection receipt, codified DESIGN.md/tokens/assets, four-surface transfer proof, and a rights-aware delivery manifest for full identity work. The one board remains a useful summary or bounded concept artifact. Source: X 2084343196994224458; Pixel Orb; Runline AI; live sites and complete local video reviewed 2026-08-12

The two Drapz posts add two refusal rules. SYNC-001 is retained because the single board shows mark variants, palette, motif, campaign application, and environmental transfer, but its only resolved destination is a booking page; the underlying product/client and reuse rights remain unverified. The portfolio reel spans Aether Studios, Radiant, Thesis, Kinect, and Quovo, so agents must decompose a montage by project and artifact class instead of blending it into one aesthetic. Drapz's current portfolio confirms that brand work is delivered as project-specific identity, guidelines, web, UI/UX, and visual systems; it does not publish a general reuse license. Source: X 2083230040133964214, 2084293211141570717; Drapz; complete local media and site review, 2026-08-12

Personal Aesthetic Skill Route

The @fleeting_land demo is a candidate pattern for personal design-aesthetic skills. The reviewed video shows a Claude skill producing a browsable design-system mirror: named aesthetic modes such as warm editorial and dark precision, principle pages like "Quiet Motion", a motion/gallery surface, and an image-led "Aesthetic Mirror." The durable rule is that a taste skill should emit reviewable artifacts and named modes, not a vague "make it my style" prompt. Keep this as a reference pattern until there is source, an install route, or repeated evidence that Kevin should encode a personal aesthetic skill beyond the existing taste-* stack. Source: X/@fleeting_land and local video review, 2026-07-03

Upstream Source Collection

These are upstream source concepts. For the active local routing logic and workflow combinations, invoke taste and load its folded reference; see Frontend and Design Skills for the underlying decision map.

Runtime Projection

npm run skill-registry projects the consolidated source skill into supported agent runtimes. Agents should resolve the taste capability through the generated registry instead of copying upstream directories by hand. Upstream updates enter through source review, Security and Review Skills, and an explicit merge into the consolidated owner.

Project-Rule Pattern

Earlier Sigil, Reticle, and Dedalus integrations used .cursor/rules/taste-enforcement.mdc to make project-specific constraints automatic. Preserve this as a deployment pattern, but verify the current project checkout before claiming the rule is installed:

• All 519 Sigil tokens consumed via var(--s-*) only (Sigil-specific overlay)
• Variance dial defaults pinned per project (Sigil: 8/6/4 marketing, 6/5/5 product)
• Banned-pattern list applied to every code block before output
• Required UI states checklist applied to interactive components

When present, the rule should cross-reference Taste Enforcement - Anti-Slop Frontend Rules and the project's own design-system instructions. The wiki registry remains the source of the shared capability; project rules are narrower enforcement overlays.

Quality Gate

Before installing or updating any taste-* skill, run the Security and Review Skills checks:

1. Brin score: curl https://api.brin.sh/repo/Leonxlnx/taste-skill - current verdict is caution, so do not import fresh upstream bodies without a manual audit.
2. Skill auditor: 6-step vetting (typosquat, permissions, dependencies, prompt injection, network, content)
3. Reputation: github.com/Leonxlnx/taste-skill - Awwwards-cited, mirrored on tasteskill.dev

The framework is content-only (no executable code, no MCP tools, no network calls), so the supply-chain risk is low - the audit focuses on prompt-injection patterns inside the SKILL.md files.

When To Extend

Add a new taste-<archetype> skill when:

• A visual style repeats across 3+ projects and the existing 4 archetypes (soft / minimalist / brutalist / GPT) don't capture it
• A workflow pattern (e.g. image-first, audit-and-fix, brand-kit) needs its own modifier skill
• A new agent surface (e.g. Anthropic skills, Codex skills, OpenCode skills) needs the framework ported

Use Agent Operations Skills to scaffold, mirror upstream to Leonxlnx/taste-skill via PR, and register in Frontend and Design Skills.

Related Tools and References

• Skills.sh - Agent Skills Registry - agent skill registry where taste-skill is published
• Design Engineering Resources - broader design-engineering tool index (Rauno, Emil Kowalski, Josh Puckett)
• Frontend and Design Skills - Kevin's adaptation with vocabulary bottleneck + archetype map
• Taste Enforcement - Anti-Slop Frontend Rules - design-page reference for the full anti-slop ruleset
• Frontend and Design Skills - routing layer for which skill fires when
• Frontend and Design Skills - Anthropics' distinctive-frontend skill (complementary, not redundant)
• Frontend and Design Skills - broader 2024-2026 design trend reference
• github.com/Leonxlnx/taste-skill - upstream source
• tasteskill.dev - published documentation site

Where Kevin Uses It

---

Timeline

• 2026-08-12 | Deep-replayed SYNC-001, the Drapz portfolio montage, and Runline through the creators' live sites. Corrected taste-brandkit from “complete world in one image” to a research → distinct directions → explicit selection → codified system → cross-surface proof → delivery workflow; added reel decomposition and expression-rights boundaries. Source: X 2083230040133964214, 2084293211141570717, 2084343196994224458; Drapz, Pixel Orb, Runline AI
• 2026-08-10 | Re-reviewed Taste through the curated design-skills bookmark; corrected the v3 ownership model to one consolidated executable taste skill with folded references, refreshed the upstream snapshot to HEAD e988add20dab0fa97d7a76781c48961c8184288e / 75,029 stars, and retained upstream as gated source evidence rather than bulk-reinstalling thirteen peer skills. Source: X/@nurijanian, 2026-07-10; Source: captured GitHub repository and README, 2026-08-10
• 2026-07-03 | Added Second Eight's Radial health-AI brand identity board as taste-brandkit evidence for technical/domain-specific identity: logo geometry, scientific textures, data coordinates, typography, UI shells, workflow cards, and product copy need to work as one system. Source: X/@secondeight_, 2026-04-09; Source: local image artifact, 2026-07-03
• 2026-07-03 | Added the jamescoder12 Gemini prompt thread as taste-brandkit prompt-structure evidence: strategy foundation, logo concepts, color, type, visual direction, voice, and brand-bible output. Kept it as a checklist for layered identity exploration, not an anti-designer thesis. Source: X/@jamescoder12, 2026-03-12
• 2026-07-03 | Added @fleeting_land's personal design-aesthetic Claude skill demo as a reference-only route: taste skills should output named modes and reviewable system pages, not only style prompts. No active skill was added because there is no public source or install route. Source: X/@fleeting_land, 2026-05-12
• 2026-07-02 | Added DrapzDZN's Radiant brand board as taste-brandkit evidence: logo mark, spacing, typography, color, usage, and application mocks. Kept it as brand-kit workflow evidence, not a new tool or proof that designers are obsolete. Source: X/@DrapzDZN, 2026-03-14; Source: local image artifact, 2026-07-02
• 2026-07-02 | Deep-reviewed X bookmark 2050697067882746209; repaired the local thumbnail, sampled the video contact sheet, resolved the self-thread to imagegen-frontend-web/SKILL.md, refreshed GitHub reputation to 54,624 stars / 3,752 forks, and recorded Brin caution for the upstream repo/skill. Source: X/@LexnLin, 2026-05-02; GitHub API and Brin, 2026-07-02
• 2026-06-30 | Deep-reviewed the imagegen-frontend-web bookmark. Promoted the durable workflow: taste-constrained visual references before UI code for high-stakes frontend pages. Updated upstream reputation/status: MIT, no tags, HEAD 06d6028b5c623016c59ce8536f578e5a1127b499, ~53.8K stars. Source: X/@LexnLin, 2026-05-02; Source: GitHub, 2026-06-30
• 2026-06-17 | Added inline provenance and timeline for the framework page; kept this as the canonical tool-level overview rather than duplicating individual taste skill docs. Source: https://github.com/Leonxlnx/taste-skill; skills/personal/taste-*
• 2026-05-04 | Created page for the upstream taste-skill framework and Kevin's local installation pattern. Source: https://github.com/Leonxlnx/taste-skill; https://www.tasteskill.dev/