Native skills vs current surfaces — what should we replace

Companion to the "Native skills + connectors" note. Native integrations aren't only additive — some of them are better-shaped versions of what we already ship. This walks each current surface and calls out replacement candidates.

Obvious replacement candidates (cost > value today)

Current surface Replacement Why
Desktop app (Electron, mac/win) PWA install + macOS Share Sheet ext + QuickLook plugin Electron wraps mdfy.app web — value ≈ web. Maintenance is heavy (build / signing / auto-update / memory). The actual mac-native value (Share Sheet, Spotlight, QuickLook) is not reachable from Electron. Drop Electron → split into three mac-native pieces. All three together are still lighter than Electron upkeep.
Chrome extension's "AI capture" mode Claude Skill + ChatGPT GPT + Gemini Extension (once all three ship) A native skill in each AI is one click in the AI's own surface — and reaches iOS / Android / desktop apps the Chrome ext can't. But the Chrome ext's GitHub README + generic web-page capture stays. Scope down to "web page capture only."

Partial replacement / depends on usage

Current surface Alternative Verdict
VS Code extension Claude Code hooks + Cursor MCP Case is weaker — if the user is already in Claude Code / Cursor, they don't open the VS Code ext separately. But the VS Code ext is an editor, not just an integration. Audit usage; consider freezing new feature work; invest in Claude Code hooks instead.
Manual AGENTS.md / CLAUDE.md / .cursor/rules wiring Claude Code hooks auto-managing the file Keep. The manual one-line recipe is the universal path that works in every tool, current and future. Hooks are an additive convenience layer for Claude Code specifically. Both.

Keep as-is

Surface Reason
MCP server IDE-tier integration (Claude Desktop, Code, Cursor). Complementary to a web-tier Skill, not redundant.
GitHub Action sync Repo / team-level source-of-truth on push. Claude Code hooks are individual + real-time — different scope. Both.
6 import routes (PDF / DOCX / Obsidian / Notion / GitHub / URL) Content ingestion paths. Orthogonal to AI integrations.
REST API Every native skill calls this. Turning it off kills everything else.

Strongest single recommendation

Drop the Electron desktop app → "mac-native triplet."

The Electron app has the highest current upkeep cost (auto-update, code signing, RAM footprint, separate GitHub Releases channel) and the lowest unique value vs. the web app. Three real mac-native scenarios users would actually use are not reachable from Electron:

  1. Spotlight / Raycast search for mdfy docs — PWA install + URL scheme handler
  2. "Save to mdfy" from Safari / Mail / Notes — macOS Share Sheet extension (Swift App Extension)
  3. Finder space-bar on a .md file → rendered preview — QuickLook plugin

The three combined are still lighter than maintaining Electron. After they ship, give the Electron app a 1-year deprecation notice and point the download page at PWA install.

Second recommendation

Migrate Chrome ext's AI-capture path to native skills incrementally.

The Chrome ext's per-AI capture (ChatGPT / Claude / Gemini chat selectors) is the most fragile piece in our stack — every UI redesign at OpenAI/Anthropic/Google forces a selector patch. Native skills don't have that brittleness. Plan: ship Claude Skill + ChatGPT GPT first, then narrow the Chrome ext to "generic web-page + GitHub README capture only."

Re-prioritised order (combining build + replace decisions)

  1. ChatGPT GPT + Actions — relieves Chrome ext AI-selector maintenance for ChatGPT users
  2. Claude Skill (web) — relieves Chrome ext AI-selector maintenance for Claude users
  3. Electron → PWA + Share Sheet + QuickLook — net upkeep reduction
  4. VS Code ext usage audit — decide freeze vs. continue based on telemetry

#1 + #2 are net-positive on day one (broader reach, lower upkeep). #3 needs the three mac-native pieces shipped before deprecation can start.