360 Secure SkillHub
360 安全技能中心

Design Engineering

核心功能是通过「研究→规划→执行→验证→优化」的迭代循环,协调子代理处理前端可视化工程任务,含技术方案选型(如Canvas/SVG/CSS抉择)、多轮优化、反馈迭代等。典型场景包括需多轮打磨的视觉任务、技术路线决策、子代理协同作业、基于反馈的实现优化。所有操作限于本地项目目录,无数据外漏,无需外部服务,输出为迭代后的前端代码及验证结果。

corbin-breton
corbin-breton
热度 65.85
最近更新 2026-05-11 17:36:41
分类:开发依赖:读写文件 / 代码执行 / 联网浏览搜索 / 内部调用大模型应用场景:代码生成 / UI 原型 / 工作流 / 研究 ...

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add design-engineering
Metadata
name
design-engineering
description
Orchestrate iterative design and frontend engineering work through research, planning, sub-agent execution, and validation loops. Use when a visual/UI task requires multiple iterations, when the first implementation needs refinement based on feedback, when choosing between competing technical approaches (Canvas vs SVG vs CSS), or when coordinating sub-agents on design-heavy work. Covers animation architecture decisions, progressive enhancement patterns, performance-aware rendering choices, and the research→plan→execute→validate workflow. Complements frontend-design (which handles aesthetics) by adding engineering discipline, iteration management, and technical decision-making.
version
1.2.0

Design Engineering

Orchestrate design-heavy frontend work that requires iteration, technical decisions, and validation. This skill is about the process — how to research, plan, build, validate, and refine — not the aesthetics (use frontend-design for that).

When to Use

  • Visual work that will take multiple iterations to get right
  • Choosing between technical approaches (Canvas vs SVG vs CSS, etc.)
  • Coordinating sub-agents on parallel design/engineering tasks
  • Refining an implementation based on user feedback
  • Any frontend work where "build it and ship" isn't enough

Scope & Safety

  • All file operations stay within the user's project directory — no absolute paths outside the project root
  • Sub-agents operate within the project scope defined by the orchestrator and must not access files outside it
  • Sub-agent outputs (file writes, research findings) are confined to the project working directory
  • No credentials or external services are required
  • Playwright (when used for screenshots) connects only to localhost dev servers — never to external URLs
  • No data is exfiltrated or transmitted to external endpoints; all work is local

The Iteration Loop

Every design-engineering task follows this cycle:

Research → Plan → Approve → Execute → Validate → Feedback → Refine

1. Research Phase

Before touching code, understand the problem space. Spawn 2-3 research sub-agents in parallel with different lenses:

  • Inspiration/reference — find examples of what we're trying to achieve
  • Technical approaches — what tools/techniques solve this class of problem
  • Constraints — performance, browser support, accessibility, progressive enhancement

Research agents should write findings to files within the project's working directory so context is preserved across iterations.

2. Plan Phase

Synthesize research into a concrete plan. Present to user for approval before executing. The plan should include:

  • Architecture decision with rationale (not just "use X" but "use X because Y fails at Z")
  • Layer/component breakdown
  • What gets removed, what gets added, what stays
  • Pre-mortem: "what would cause this to fail?"
  • Estimated sub-agent tasks
  • Horizon estimate: Count estimated steps across all sub-agents. If > 40 steps total, decompose into sub-phases of ≤ 20 steps. If 20–40 steps, insert a midpoint quality checkpoint (screenshot + integration check before proceeding). Under 20 steps, execute directly. This prevents quality collapse on complex design tasks.

3. Execute Phase

Dispatch sub-agents with focused, context-minimal tasks. Key rules in references/subagent-patterns.md.

4. Validate Phase

After sub-agents complete, the orchestrator MUST validate. Build check is necessary but not sufficient. Check integration points — see references/validation-checklist.md.

5. Feedback → Refine

Ship to user for review. Expect 2-5 iterations on visual work. Each iteration:

  1. Screenshot the live result (use Playwright if available)
  2. Identify specific issues from feedback
  3. Make targeted fixes (don't rebuild from scratch each time)
  4. Validate and redeploy

Technical Decision Framework

When choosing between rendering approaches, read references/rendering-decisions.md. Quick heuristic:

Need Use
Static decorative pattern CSS background-image with SVG data URI
<100 authored animated elements Inline SVG + CSS animations
Procedural generation, >100 elements, full-page coverage Canvas 2D
3D, heavy particle systems, post-processing Three.js/WebGL (last resort — heavy)

Progressive Enhancement Stack

Every visual enhancement must degrade gracefully:

  1. CSS baseline (always works, no JS)
  2. JS-enhanced layer fades in on top
  3. prefers-reduced-motion → skip animations entirely
  4. Low-end device detection → reduce complexity
  5. Light/dark theme awareness

Glass-Panel Pattern

For content floating over animated backgrounds:

.card {
  background: color-mix(in srgb, var(--bg-card) 50-60%, transparent);
  backdrop-filter: blur(8-12px);
  -webkit-backdrop-filter: blur(8-12px);
}

Lets animation show through while keeping text readable. Adjust blur and opacity based on background intensity.

References

  • references/subagent-patterns.md — How to dispatch and validate sub-agent work
  • references/validation-checklist.md — Post-execution checks that catch integration bugs
  • references/rendering-decisions.md — Canvas vs SVG vs CSS decision guide with production lessons