A Claude Code plugin that catches AI-generated design defaults before they ship.
Kern routes every design through a persona system, picks a varied subset of anti-patterns per run (so two runs never produce the same critique), scores output on a 0-100 sameness scale, and fixes copy, color, layout, and interaction before the code lands.
Runs a team of specialist agents that plan, design, implement, and review collaboratively.
AI design tools converge on the same visual vocabulary. Purple/blue gradient, centered hero, Inter font, three identical feature cards, "Get Started" CTA. A developer tool and a consumer app look like the same product built on the same template.
Kern fixes this with persona-first generation and a sameness score that quantifies how close a design is to the AI default. It runs inside Claude Code, auditing the code Claude generates for you.
Add the raymondware marketplace, then install Kern:
/plugin marketplace add raymondware/kern
/plugin install kern@raymondware
That's it. Restart Claude Code and the /kern:* commands are available.
# Generate a component from scratch
/kern:design hero section for a CI/CD observability platform
# Audit existing code for AI-default patterns
/kern:audit
# (paste TSX when prompted)
# Make an existing design less AI-looking, optionally against a benchmark
/kern:differentiate [--reference URL | recipe:name@version]
# Generate design specs without code, optionally with a recipe or reference
/kern:plan pricing page for a developer API product| Command | Does |
|---|---|
/kern:design <desc> |
Full pipeline: persona detection, specialists, implementation, review |
/kern:audit |
Anti-pattern scan with sameness score against existing code |
/kern:differentiate [--reference ...] |
Identify AI fingerprints and apply targeted fixes against an optional benchmark |
/kern:plan <desc> |
Design specs only (typography, color, layout, motion) -- no code |
/kern:review |
Full review pass: critic ensemble, copy editor, accessibility audit |
/kern:compare <a> <b> [--reference ...] |
Side-by-side comparison with similarity score and optional benchmark adaptation |
/kern:polish |
Rewrite a component against Kern's standards, show diff |
/kern:copy |
Microcopy audit: button labels, errors, empty states, CTAs |
/kern:research |
Scan design communities for new anti-patterns |
/kern:draw <desc> |
Run the anti-pattern selector standalone (rotation verification) |
Kern picks a persona from your description, then loads that persona's rules for font, color, layout, and copy. Five personas:
| Persona | Register |
|---|---|
developer-tool |
Restrained, dense, function-first. Linear/Vercel register. |
consumer-saas |
Warm, spacious, approachable. Things 3 / Cron register. |
creative-tool |
Bold, editorial, personality-forward. Figma / iA Writer register. |
b2b-enterprise |
Conservative, trust-building, data-forward. Stripe / Ramp register. |
e-commerce |
Product-first, conversion-focused. Allbirds / Ridge register. |
Every run picks a different subset of anti-patterns. Kern reads the audit log (state/draws.jsonl) and excludes the previous run's subset. Two runs on the same prompt produce different critiques. No template drift.
Each run draws 12-18 patterns from a pool of 69, weighted by surface (landing page, dashboard, pricing), persona, industry, and audience. The selection is logged to state/draws.jsonl before any agent runs.
Four critics run in parallel (design, hierarchy, interaction, microcopy), plus a copy editor and accessibility auditor. The synthesizer merges their output, removes duplicates, and computes a consensus sameness score. The post-synthesis design-gauntlet separately checks task clarity, hierarchy, distinctiveness, content authenticity, interaction quality, responsive integrity, reference adaptation, and baseline regression when applicable. It can trigger at most two targeted cycles and reports unresolved risks when it cannot pass.
Every design gets a 0-100 score on AI-default similarity:
| Score | Meaning |
|---|---|
| 0-20 | Distinctive. Clear decisions made. Ship. |
| 21-40 | Has character. Some defaults showing. |
| 41-60 | Generic. Could be any AI-generated SaaS site. |
| 61-80 | Default AI aesthetic. Multiple tool fingerprints present. |
| 81-100 | Template unmodified. Do not ship. |
Gate threshold: 40 for Kern-produced output, 60 for external audits.
Kern ships with 69 anti-patterns across three base files and a research-sourced directory:
| File | Patterns |
|---|---|
anti-patterns/visual.md |
Gradient slop, nested card soup, Inter default, centered glowing hero, AI palettes, drop shadows, inverted dark mode, neon without context |
anti-patterns/copy.md |
Marketing verbs, hedging, filler openers, self-descriptors, em-dashes, passive voice, vague headlines, broken microcopy |
anti-patterns/interaction.md |
Dishonest CTAs, loading as marketing, motivational empty states, bouncy interactions, modal fatigue, toast spam, dark patterns |
anti-patterns/sourced-from-research/ |
Community-sourced patterns with citations |
| File | Purpose |
|---|---|
skills/kern/references/fonts.md |
Font stacks by context: display, body, mono, serif |
skills/kern/references/tokens.md |
Spacing, radius, type scale, color philosophy |
skills/kern/references/exemplars.md |
Sites that get it right and why |
skills/kern/references/dribbble-refs.md |
Reference shots by category |
skills/kern/references/ai-fingerprints.md |
Tool-specific fingerprints for /kern:differentiate |
skills/kern/references/personas/ |
Five persona files with font, color, layout, copy rules |
skills/kern/references/orchestration/reference-led-design.md |
Reference benchmark, adaptation, provenance, and divergence contract |
skills/kern/references/style-recipes.md |
Versioned and scoped style recipe schema used as an input to design, compare, or differentiate |
Kern runs a team of specialized agents:
typography-specialist color-specialist layout-specialist motion-specialist
component-architect component-implementer style-implementer accessibility-implementer
design-critic hierarchy-critic interaction-critic microcopy-critic copy-editor accessibility-auditor responsive-validator
conductor anti-pattern-selector critique-synthesizer
See CLAUDE.md for the contributor guide. Agents live flat in agents/. Anti-patterns in anti-patterns/. Commands in commands/.
MIT. See LICENSE.