design-garden

agent
Security Audit
Warn
Health Warn
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 6 GitHub stars
Code Pass
  • Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

The design skill that grows - a universal, self-improving design skill for AI agents (UI, posters, icons, logos, slides, data-viz, video, illustration) | 会成长的设计技能:面向 AI 智能体的通用自进化设计技能(UI、海报、图标、Logo、幻灯片、数据可视化、视频、插画) | 🌱 新手入门站 · start here ↓

README.md

🌐 English · 简体中文 · 繁體中文 · 日本語 · 한국어 · Español · Português · Français · Deutsch · Русский

design-garden — the design skill that grows · 会成长的设计技能

🌱 design-garden

Plant a design skill. Watch it grow.
「种下一个设计技能,然后看着它长大。」

License: MIT
GitHub stars
CI
styles
domains
learnings

A universal, self-improving design skill for AI coding agents — covering anything aesthetic: UI, posters, banners, icons, logos, brand identity, slides, data visualization, video & motion, illustration, typography, and design critique.

▶ 60-second demo · Knowledge base · What it makes · Why it's different · What's inside · Install · Contribute

🌱 New to AI tools? No-code install, plain-language usage, and zero-skill ways to help: Start Here (中文:新手指南) — and the beginner site explains every GitHub button, no code needed.

What it makes

The portfolio — 10 works across 9 craft domains. Every piece was made by the skill running its own workflow: a three-direction DEAL before the build (evidence kept), lint.mjs + rendered GATE before delivery, and a learning note written back afterwards. Each work below is labeled with its domain, format, and the model that produced it.

Bike repair shop landing page hero Apiary quarterly member report deck, cover slide Bike-share monthly rider report, lead chart
Landing page · ui-web · 1440 + 375
GLM-5.3-Flash · source · mobile · full page
Member-report deck · slides · 7 × 1920×1080
GLM-5.3-Flash · overview wall · harvest chart · varroa chart
Rider report · data-viz · 1200px data page
GLM-5.3-Flash · source · full render
Jazz cafe opening night poster Night-running beginner carousel cover, Chinese Plant-care icon set verification sheet
Opening-night poster · poster-banner · 1080×1350
GLM-5.3-Flash · source · direction 2
Night-run carousel · social-carousel · 7 × 1080×1440, 中文
GLM-5.3-Flash · source · route page · closing CTA
Plant-care icon set · icon · 8 glyphs, 16/24/48 ladder
GLM-5.3-Flash · source · SVGs
Peso budgeting onboarding prototype, welcome screen Percolate coffee-cart logo system and mini brand guide board Sleep-debt long-scroll infographic cover, Chinese
Onboarding prototype · ui-mobile · tap-through, 6 states, 375×812
GLM-5.3-Flash · source · failed state · empty home · landscape
Logo system + guide · logo-brand · 4 variants × 4 colorways
GLM-5.3-Flash · board · full board · DEAL sheet
Sleep-debt longform · infographic-long · 750×9054, 中文
GLM-5.3-Flash · source · week bars · cumulative

DEAL sheet — three structurally different first screens for the deck brief, side by side

▲ The DEAL gate as a picture — three structurally different first screens built and compared before the deck; the report dealt its lead figure and the carousel its cover the same way. Evidence and the gate's caught-and-fixed list: examples/README.md.

▶ Watch the 60-second loop — one real run narrated start to finish: the brief, the three-direction deal, the gate's actual RED catches (CTA hover at 3.3:1, nav at 2.6:1, an 11.5px label), the fixes, and the learning note written back. Every frame is a file from this repo. Or open demo/demo.html from a clone.

One brief, eighteen styles — the style gallery

18 style specimens of the same fern-landing brief — 18 different skins on one skeleton

▲ 18 style specimens on one brief — same copy, same layout skeleton, 18 different skins. Every image in this README was made by the skill itself. Browse the specimens

The stills are real outputs — every piece went through the full workflow, critique gate included. Generated with GLM-5.3-Flash — per-example model attribution in examples/README.md. Run the same brief on another model and PR your result.

Most agent "design skills" are frozen prompt packs: strong on day one, stale on day thirty, and always one author's taste. design-garden is built to be fed — by your daily design work, and by the community. It doesn't just execute design tasks; it learns from every task it runs.


Why it's different

design-garden Frozen prompt packs Single-domain skills
Coverage All 9 design domains + 18 styles Varies One domain, deep
Growth Learns from every task + community PRs Never changes Manual forks
Quality control Eval-gated promotion + CI injection scan None Rarely
Beginner paths Designed-in (vote, report, translate) None None
Bloat control ≤6 files loaded per task, line budgets Grows forever N/A

The four growth loops

flowchart TB
    A["① Runtime self-learning<br/>every task appends a structured entry to learnings/ —<br/>the skill gets better for YOU with every use, no reinstall"]
    B["② Community contributions<br/>knowledge = small modular files.<br/>One PR = one new style recipe / technique / case"]
    C["③ Eval-gated promotion (anti-bloat)<br/>new knowledge lands as experimental, loads only on explicit match,<br/>and becomes proven only after winning a blind A/B eval"]
    D["④ Versioned learning<br/>every release's CHANGELOG says what the garden LEARNED"]
    A -->|"generalizable entries get PR'd upstream"| B
    B --> C
    C --> D
    D -.->|"feeds the next task"| A

What's inside

SKILL.md                  router + 8-step workflow (≤500 lines — the stable core)
knowledge/
  core/                   universal modules: principles · color · typography · layout · motion · anti-slop
  domains/                9 domains: ui-web · ui-mobile · poster-banner · icon · logo-brand · slides · data-viz · video-motion · illustration-art
  styles/                 18 style recipes (swiss, bauhaus, neo-brutalism, glassmorphism, dark-tech, y2k, japandi, memphis…) — one small file each
  case-studies/           before/after with reasoning
learnings/                the skill's field notes — starts seeded, grows with every task
evals/                    benchmark briefs with deliberate traps + rubric + blind A/B harness
scripts/                  validate_knowledge.mjs (CI gate) · ab_harness.mjs (blind eval pages)
                          shoot.mjs · lint.mjs · ledger.mjs · directions.mjs (delivery tools)
                          gallery.mjs (18-style gallery builder, repo maintenance)

Every knowledge file carries frontmatter (id / kind / applies / status / source / updated). experimental knowledge loads only on explicit match; proven loads by default. The router reads knowledge/INDEX.md and never loads more than ~6 files per task — growing without bloating.

Four delivery tools ship in scripts/: shoot.mjs renders at exact sizes and builds a contact sheet, lint.mjs runs the mechanical half of the GATE (contrast, floors, transition: all, motion escapes, the Never list), ledger.mjs remembers the look of your last deliveries so the next one does not repeat it, and directions.mjs renders 2–3 real 1440×900 first screens and composes the DEAL sheet. None of them blocks: Playwright is used when installed, a local Chrome/Edge over the DevTools Protocol otherwise, and when neither exists the script exits 2 and says what it could not verify. gallery.mjs is repo maintenance — it re-renders the 18 style specimens and writes examples/styles/index.html + overview.png.

The workflow itself is a distillation of the strongest process ideas in agent design practice: a three-direction deal (2–3 structurally distinct art directions, each shown as a real 1440×900 first screen before the full build), the five form questions, a critique gate that fixes only the 3 most damaging issues per round, and a canonical anti-slop list merged from three independently-convergent sources. See SOURCES.md.

Install

One command, 20+ agents. The skills CLI (the open agent-skills registry) installs design-garden into Claude Code, Codex, Cursor, Windsurf, Gemini CLI, OpenCode, ZCode and other SKILL.md loaders — it detects which agents you have and offers them as a picker:

npx skills add iifzx/design-garden      # pick your agent interactively
npx skills add iifzx/design-garden -g   # user-level, all projects

Other routes, same destination:

# plugin marketplace (Claude Code)
/plugin marketplace add iifzx/design-garden
/plugin install design-garden@design-garden

# via npm/github (cross-platform installer)
npm i -g github:iifzx/design-garden && design-garden

# classic git clone (user-level, all projects)
git clone https://github.com/iifzx/design-garden.git ~/.agents/skills/design-garden

# DeepSeek Harness (dsh) — skill roots are non-recursive, so clone to a skill root
git clone https://github.com/iifzx/design-garden.git ~/.dsh/skills/design-garden

# or project-level
git clone https://github.com/iifzx/design-garden.git <your-project>/.zcode/skills/design-garden

Every route serves the same vetted tree: CI runs the knowledge validator and an injection scan on every commit, and community knowledge stays experimental (loads only on explicit match) until it wins a blind A/B eval — the install never contains unreviewed load-by-default rules.

DeepSeek Harness note: dsh reads the same Agent Skills standard (SKILL.md dirs, non-recursive discovery under ~/.dsh/skills or ~/.agents/skills). The skill declares whenToUse for dsh's trigger metadata.

ZCode desktop: Plugin Marketplace → Add → paste https://github.com/iifzx/design-garden. The marketplace layout follows the Claude Code plugin convention; if your CLI's loader expects different nesting, the git-clone path always works — report it as an issue so we can support your loader too.

No build step. The scripts are Node with stdlib only; shoot.mjs gets full-page and per-screen support from Playwright when it is installed, and otherwise drives a local Chrome/Edge over the DevTools Protocol. Nothing is required beyond Node.

Use it

Just design. The skill triggers on anything aesthetic:

  • "Make a landing page for my ceramics studio" → deals 3 art directions, executes the strongest, critiques, delivers, records what it learned
  • "Design an icon set for a plant-care app" → keyline grids, optical corrections, silhouette tests
  • "This poster looks off — review it" → critique-first mode, runs the domain checklist, gives fixes ranked by damage
  • "品牌海报 + 一套 PPT 模板" → routes both domains, learns both

Before a multi-canvas build, directions.mjs turns 2–3 directions into one labeled first-screen sheet; after delivery, the three tools close the loop: shoot.mjs shows the page at 1440 and 375, lint.mjs lists floor violations with line numbers, ledger.mjs add records the look so the next project starts somewhere else.

For contributors

The garden grows on small PRs. Add a style recipe (~1 hour), report a learning from your own use (minutes), or run an eval to promote an experimental rule to proven. Everything is in CONTRIBUTING.md — the short version:

  1. Fork, add one small file following the schema (or one learning entry)
  2. node scripts/validate_knowledge.mjs must pass (CI enforces it)
  3. Open the PR — CHANGELOG will say "the garden learned X"

Community knowledge starts experimental (loads only on explicit match) and becomes proven by winning a blind A/B eval — so growth is verified, not just accumulated.


Translations

English is the primary version — the skill content (SKILL.md, knowledge files) and trigger vocabulary stay English for model reliability. Community translations of the README live alongside it:

The v0.1.5 translations were drafted by GLM-5.3-Flash — improvements from native speakers are very welcome (see CONTRIBUTING.md).

Contributors

This garden lives on contributions, recognized via all-contributors (.all-contributorsrc):

@iifzx 💼 📖 🎨 🚧 🌍 — code, docs, design, maintenance, translations

Feed it and add yourself: comment @all-contributors please add @<you> for <contributions> on any issue, or PR the table above.

Star history

Star History Chart

License & credits

MIT — see LICENSE. Seeded by distilling 17 design skills and the shared craft canon; full credit in SOURCES.md.

🌱 Gardens are judged by what they grow.

Reviews (0)

No results found