由于这篇文章太伟大了，所以我把它变成了一个 Agent Skill。 大家可以使用自己的 Coding Agent 安装一下这个 Skill，这样就可以用「最佳实践」来轻松地重构或者开发一个既容易跨平台、又极其接近 Native 性能的桌面端应用。 https://github.com/yetone/native-feel-skill…

yetone/native-feel-skill

Source: https://github.com/yetone/native-feel-skill

Install the native-feel skill for me:

1. Clone https://github.com/yetone/native-feel-skill into your
   user-level skills directory as `native-feel-cross-platform-desktop/`
   (whatever path your agent reads skills from on this machine — e.g.
   `~/.claude/skills/` for Claude Code).
2. Verify that SKILL.md plus the references/ and checklists/ subdirs
   are present.
3. Confirm the install path when done.

Use cases

Once installed, the skill activates automatically whenever the conversation touches cross-platform desktop architecture, WebView quirks, or making a UI feel native. Two common ways to use it:

Case 1 — Refactor an existing app to feel more native

You have an Electron / Tauri / web-wrapped app, and a designer (or your own eye) tells you it feels web-y. You want to know where to start and how far you can take it without a full rewrite.

You    ❯ My Electron app feels web-y on macOS — the designer says it
         doesn't pass for a real Mac app. Where do I start?

Agent  ❯ Quickest path is the 75-item native-feel audit
         (checklists/ship-readiness.md). Most Electron apps fail
         the same 6–8 items:
           - cursor:pointer on rows (item 21)
           - web-style modal overlays instead of native sheets (19)
           - hardcoded brand accent instead of system accent (33)
           - page fade-transitions (40)
           - opaque window background instead of platform material (31)
           - WebKit context menu still firing (23)
         Each is a 5-to-30-minute fix; together they move you most
         of the way to "passes the 30-second skeptical user test."

         Bigger question first: how far do you need to go? If "good
         enough" means passing that 30-second test, Electron + this
         audit gets you 80% there without rewriting. If you want
         "indistinguishable from a Mac app" — Liquid Glass materials,
         native popovers, sub-200ms launch — Electron has a ceiling
         you'll hit. Want me to run the decision tree first?

Case 2 — Build a cross-platform native-feel app from scratch

Greenfield project, must run on macOS + Windows, and the native feel is the point — a launcher, a productivity tool, an AI workspace your users live in all day.

You    ❯ Starting a new desktop app — macOS + Windows, native feel
         is the whole pitch. What architecture should I use?

Agent  ❯ Before architecture, the decision tree
         (checklists/decision-tree.md) rules this stack out for
         several common shapes — quick check first:
           - Cold-start budget?  (<100ms → build native instead)
           - Memory floor?       (<150MB → build native instead)
           - Extension ecosystem? (yes → need the Node layer)
           - Runway?             (<3mo → start on Electron)

         If you pass, the four-layer architecture is:

           ┌──────────────────────────────────────────────────┐
           │ Native shell      Swift/AppKit  +  C#/WPF        │
           │       ↕   one IPC schema, codegen for all sides  │
           │ System WebView    WKWebView    +  WebView2       │
           │   running React + TypeScript, shared 1:1         │
           │       ↕                                          │
           │ Node backend      single long-lived process      │
           │       ↕                                          │
           │ Rust core         UniFFI-bridged, sharable       │
           │                   with iOS and server            │
           └──────────────────────────────────────────────────┘

         The expensive-to-get-wrong decisions, in order:
           1. The IPC contract (one schema, codegen). Get this wrong
              and you rebuild it later under duress.
           2. WebView survival flags (occlusion off, alpha-prewarm,
              _doAfterNextPresentationUpdate, font prewarming).
           3. Memory hygiene (lazy-load extensions, tear down
              secondary windows aggressively).

         Want to start with the seam — where to draw the cross-
         platform line — or with the IPC contract, since that's the
         load-bearing piece?

Other things the skill is good for

• “Why is my WKWebView flickering when I hide and re-show it?” → walks you through references/03-webview-survival.md (most likely A.1 throttling or A.2 startup flicker).
• “How should typed IPC work across Rust, Swift, and TypeScript?” → the UniFFI-based pattern in references/04-ipc-contract.md, with the exact Coordinator/EventHandler shape Raycast Beta ships.
• “My app is at 450 MB resident, is that bad?” → the six common Activity-Monitor mistakes and what to actually measure, in references/05-memory-truths.md.
• “Is my designer’s spec ‘native enough’?” → the 70+ item conventions audit in references/06-native-conventions.md.

“We’re not a web app with some native hooks sprinkled on top. We’re a native app that uses web for its UI.” — Raycast

What this is

A reference for architects, tech leads, and engineers who must build a desktop app that:

• runs on macOS + Windows (optionally Linux) from a single UI codebase,
• launches in under 500 ms and stays under 500 MB resident,
• is indistinguishable from a native app to its users (no cursor: pointer tell-tales, no white-flash on launch, no WebKit context menu, no smooth-scroll JS),
• supports a plugin/extension ecosystem in TypeScript,
• can share performance-critical code with iOS and a server backend.

This is the four-layer architecture: native shell → system WebView (WKWebView/WebView2) → Node backend → Rust core, wired together with a single typed IPC schema that generates clients for every runtime.

What this is not

• Not for single-OS apps (just build native).
• Not for Electron-style “good enough” apps (the polish budget here is 5–10× higher).
• Not for apps with strict <150 MB or <100 ms cold-start budgets (the floor is real).
• Not for games, document editors, or media players.

Run checklists/decision-tree.md to find out if this architecture is even right for your project. It rules itself out for several common cases — saying so directly is more useful than over-fitting advice.

Layout

native-feel-skill/
├── SKILL.md                                # entry point for the agent
├── references/
│   ├── 01-philosophy.md                    # 8 tenets that drive every decision
│   ├── 02-architecture.md                  # the four-layer architecture
│   ├── 03-webview-survival.md              # WebKit/WebView2 quirks + fixes (the goldmine)
│   ├── 04-ipc-contract.md                  # typed IPC across Rust/Swift/C#/TS
│   ├── 05-memory-truths.md                 # why Activity Monitor lies
│   ├── 06-native-conventions.md            # 70+ items the native-feel audit checks
│   └── 07-evidence-raycast.md              # what a reverse-eng. of Raycast Beta shows
└── checklists/
    ├── decision-tree.md                    # should you use this architecture?
    └── ship-readiness.md                   # 75-item launch audit

Philosophy

The central tension this architecture resolves: how can a desktop app deliver convenient cross-platform development AND near-native performance, when those goals usually pull against each other? Eight tenets name the structural moves:

1. Place the seam at the rendering surface — share above the WebView, diverge below it; this is the only altitude where both DX and native feel survive.
2. One schema, many languages — pay the polyglot tax once at the declaration, never at the call site.
3. Adopt the platform; don’t compete with it — the OS draws blur, scrolling, materials, and dark mode better than you can.
4. Performance is a property of perception — what the user feels, not what Activity Monitor reports.
5. The short iteration loop is the product — 200 ms hot reload vs 30 s native rebuild is a 150× compounding advantage.
6. Cross boundaries intentionally — IPC has a cost; design every crossing as async, batched, schema-typed.
7. Identity is muscle memory — the hotkey, the rank order, the verbs are the app; everything else is implementation.
8. Separate baseline from margin — the WebView+Node floor is rented; only your dirty pages are yours to optimize.

Read references/01-philosophy.md first. Everything else is consequence.

About Agent Skills

Agent Skills are the emerging standard for packaging domain knowledge that any compatible agent (Claude Code, the Claude Agent SDK, or other Agent-Skill-aware runtimes) can discover and load. Once installed via the prompt at the top of this README, the skill activates automatically when the agent’s conversation touches cross-platform desktop architecture, WebView quirks, or Raycast-style apps — the trigger conditions are declared in SKILL.md’s frontmatter.

Sources

• Raycast’s public technical post: A Technical Deep Dive into the New Raycast
• Reverse engineering of Raycast Beta.app v0.60.0 (macOS 26+ build, Xcode 17, arm64) — see references/07-evidence-raycast.md for what was found and how.

License

MIT — see LICENSE.

Credits

Authored as an Agent Skill. The architecture this skill describes is Raycast’s; the philosophy is the author’s synthesis; the evidence is from the shipping app.