# garden-skills packages taste and process for AI coding agents > Source: > Published: 2026-07-29 15:04:56+00:00 ConardLi's garden-skills makes a specific bet: what holds AI coding agents back is not raw capability but taste and process discipline. Each skill in the collection ships as a full production surface with opinionated defaults and hard stop points, not a loose prompt you paste and hope works. garden-skills is a curated set of production-ready Agent Skills for Claude Code, Cursor, Codex, and other AI coding agents. It is MIT-licensed, follows the SKILL.md spec, and offers five ways to install (a `skills` CLI via npx, the Claude Code plugin marketplace, pinned `.zip` releases, manual copy, or a git submodule). The badge counts five skills. The two the README documents in depth show what the author is actually after, and it is not just "more features." `web-design-engineer` exists to move AI-generated web artifacts from merely functional to deliberate. Instead of treating the agent as an autocomplete for markup, it casts the agent as a design engineer: understand the product context first, declare a design system, show an early v0, build the full experience, then verify it. The interesting part is how much judgment the skill hard-codes. It converts a brief into a five-dial Design Read covering variance, motion, density, asset dependence, and brand fidelity. It distinguishes three redesign modes, extension, preserve, and overhaul, before it touches an existing product. It carries an anti-cliche blocklist to steer away from the generic UI patterns agents default to. And it ships a Design Direction Advisor with six differentiated schools plus 25 anchored style recipes, named after real references like Linear, Aesop, Bloomberg, Stripe Press, and Mid-Century. Each recipe comes with a concrete palette, typography, signature moves, and anti-patterns, ready to paste into the design-system declaration. Underneath sit practical rules for inline React plus Babel, CSS tokens, `oklch()` color, container queries, and reduced-motion handling, with browser acceptance testing that runs only when you explicitly ask for it. `web-video-presentation` turns scripts, articles, lessons, and talks into click-driven 16:9 web presentations that can be screen-recorded as videos. It scaffolds a Vite plus React plus TypeScript project that behaves like a video production surface: a fixed 1920x1080 stage that scales to the viewport for stable recording, a click or keyboard driven `(chapter, step)` cursor with one narration beat per visual step, and hidden hover-only progress controls so the stage stays clean on camera. Again the discipline is the point. The workflow turns raw articles into narration scripts, maps script beats to full-screen scenes, and pauses at hard collaboration checkpoints for the script, the theme, the outline, the implementation mode, and optional audio. Nothing runs ahead of your approval. It includes 23 built-in themes, each with its own design signature spanning editorial, terminal, engineering, and Swiss International looks. Audio is pluggable: a provider-agnostic runner ships two built-in providers (MiniMax `mmx-cli` and OpenAI TTS via curl) and gives you a contract plus ready-to-paste snippets for ElevenLabs, edge-tts, Azure, Google Cloud, and the macOS `say` command. The gallery lists two more skills by name: `gpt-image-2` for image generation and prompting, and `beautiful-article` for turning any source into a formatted article. The repo description and topics also point to a knowledge-retrieval, RAG-oriented skill, which rounds the count to five. The README does not document those three in the same depth, so their internals are worth checking in each skill's own SKILL.md before you rely on them. What ties the documented skills together is a consistent instinct: encode the checkpoints, the named recipes, and the anti-patterns that separate competent output from output a person actually chose. A five-dial Design Read and an anti-cliche blocklist are opinions about what good looks like. Pausing a video build at five named checkpoints is an opinion about how humans and agents should divide the work. That framing is the reason to watch garden-skills rather than any single skill in it. Most agent skill libraries expand what a model can do. This one is mostly about constraining how it does the work, so the result reads as chosen rather than generated. If that bet holds, the value of a skill collection shifts from capability coverage toward the quality of the defaults it hard-codes. **GitHub:** [https://github.com/ConardLi/garden-skills](https://github.com/ConardLi/garden-skills) *Curated by Agent Palisade — practical AI for small and mid-sized businesses.*