MacStories founder Federico Viticci just dropped Shortcuts Playground, a free and open-source plugin for Claude Code and Codex that generates real Apple Shortcuts from natural language prompts. After six months of daily iteration with both AI agents — reverse-engineering the entire Shortcuts action graph in the process — Viticci has built something genuinely wild: point your desktop agent at his GitHub repo, type a sentence describing what you want, and walk away. A few minutes later, you'll come back to a signed .shortcut file sitting in Finder, ready to import into the Shortcuts app on any of your Apple devices.

How It Actually Works

Under the hood, this is clever as hell. Shortcuts have always been XML files under the hood — they just get encrypted into Apple's proprietary .shortcut format after signing. The plugin bundles extensive documentation that teaches an LLM how every Shortcuts action works, what syntax it expects, and how actions connect to each other. When an agent generates a shortcut draft as XML, a validation hook fires what Viticci calls a Craig Loop — his take on the popular Ralph coding loop technique. The agent runs a Python validator over and over until the shortcut passes syntax checks and Apple signs it cleanly with macOS's built-in shortcuts CLI. No runtime access to your Shortcuts database required.

Model Matters — A Lot

Shortcuts Playground doesn't perform equally across all models. Viticci recommends Claude Opus 4.6 or 4.7 at Medium or High reasoning level, or GPT-5.5 also at Medium or High. His sweet spot is Opus 4.7 with Fast mode at High reasoning — fast enough to be practical while still hitting the complexity needed for advanced shortcuts. The plugin ships as a native extension in Claude Code, supporting dedicated /build and /remix slash commands that spin up sub-agents for creating new shortcuts or remixing existing ones from exported XML. Codex gets its own version too, though Viticci notes that Anthropic's more mature plugin architecture currently gives Claude Code the edge.

Six Months of Trial and Error

Viticci started with an open-source skill by GitHub user drewocarr that attempted the same thing but produced unreliable results — hallucinated actions, missing variables, broken connections. He fed it to Codex along with his own Shortcuts database (analyzed from ~/Library/Shortcuts/Shortcuts.sqlite), modernizing the action catalog from roughly 1,000 entries to around 2,000. Then came hundreds of generated shortcuts, each one tested and critiqued. When something broke, Viticci attached the output and screenshots back to Codex and asked it to diagnose the failure. Over half a year, both agents taught each other patterns, best practices, and edge cases that nobody had documented publicly before.

The Craig Loop: Self-Correcting Shortcut Generation

The biggest breakthrough wasn't the documentation — it was validation. Early builds still produced shortcuts with empty parameters and orphaned actions. Viticci borrowed a page from the agentic coding world and created a loop system where the agent keeps re-running its own output through a validator script until everything checks out. This adds latency but dramatically improves reliability. In his testing, Viticci estimates Shortcuts Playground hits about 90% accuracy on first attempts — good enough to be useful, but he stresses you should always open generated shortcuts in the app and verify them yourself.

Real-World Use Cases That Go Deep

On the simple end, you can ask for a Hello World shortcut or something that grabs your five most recent screenshots and sends them over iMessage. But because Shortcuts Playground runs on your Mac with full CLI access, it gets much more serious. Viticci used it to build shortcuts that list all running cron jobs on his machine, interact with the Notion API by appending text and images to daily notes, pull tasks from Todoist and let him reschedule them with a date picker, record audio, transcribe it via the Gemini API, and paste the result to clipboard. He even made one that controls OpenClaw directly from Shortcuts.

Apple Will Sherlock This — And That's Fine

"If I was able to make something that generates shortcuts with a prompt, imagine what Apple could do," Viticci writes. He's not worried about his work being obsoleted by a native Apple Intelligence feature in iOS 27. In fact, he hopes it happens. His philosophy is straightforward: he's spent nearly twelve years teaching people to build automations the hard way, and now that AI can handle most of the heavy lifting, his role shifts from builder to orchestrator — someone who teaches readers how to leverage agents rather than manually construct every action themselves.

Key Takeaways

  • Shortcuts Playground is live on GitHub, free and open source for both Claude Code and Codex
  • Best results with Opus 4.7 or GPT-5.5 at High reasoning; Fast mode works fine for most shortcuts
  • The Craig Loop validation system catches syntax errors automatically before signing .shortcut files
  • Agents can research and test APIs (Notion, Todoist, Gemini) independently before generating a shortcut based on real responses
  • 100 pre-generated, Viticci-verified shortcuts included in the redesigned MacStories Shortcuts Archive

The Bottom Line

This is what it looks like when someone with deep domain expertise weaponizes AI agents against their own workflow — and then gives it away for free. Six months of reverse-engineering Apple's undocumented shortcut format, distilled into a plugin that anyone can install in minutes. Whether Apple ships a native version next month or Viticci's tool becomes the de facto standard in the meantime, this is exactly how automation should have worked all along: describe what you want, and let the machine figure out the rest.