MacStories creator Federico Viticci recently launched Shortcuts Playground, a free, open-source plugin for Claude Code and Codex that takes a plain-English description and returns a signed .shortcut file ready to import into Apple's Shortcuts app. No drag-and-drop builder.
No hunting through action menus. The catch is what makes this launch different from every earlier attempt: the plugin pairs generation with a looping Python validator that keeps correcting output until it passes structural checks. That detail matters more than the natural-language interface itself.
Earlier tools produced plausible-looking output that broke on import. Shortcuts Playground's fix for that problem, and what two independent open-source projects now confirm as the right architecture for this class of AI tooling, is the more important story here.
What launched: prompt in, Apple Shortcuts .shortcut file out
Describe any automation in Claude Code or Codex, and the plugin designs the action sequence, wires the variables, assigns an icon, and packages the result as a signed binary. Shortcuts have always been XML under the hood; Shortcuts Playground is a natural-language generator for Apple's .shortcut file format, covering classic Workflow actions and modern AppIntents across iOS, iPadOS, and macOS.
The workflow extends beyond one-shot creation. Users can resume a prior session to refine an existing shortcut, or point the agent at an existing XML file and describe what to change. Generator and editor in one.
Alongside the plugin, Viticci relaunched the MacStories Shortcuts Archive with 100 shortcuts generated entirely by Shortcuts Playground and verified before publication. The plugin is free and open source, with code and documentation available on GitHub.
Who this is actually for and what setup it requires
Shortcuts Playground runs inside Claude Code or Codex. Neither is a consumer product with a simple install button, so nontechnical users will face platform setup before they reach any prompt interface. Whether that access carries usage costs is not clarified in available documentation.
Once past that step, the barrier to building shortcuts drops considerably. Viticci describes the tool as capable of letting someone "will an idea into existence" without any prior Shortcuts knowledge — no understanding of action types, variable scoping, or the visual builder required. The friction it removes is shortcut-building friction. Platform-access friction is a separate problem that this tool doesn't solve.
One practical note on security: signed shortcuts produced by tools like this can run with broad system access once imported. ShortcutForge, a separate open-source project, notes that importing signed shortcuts requires enabling Private Sharing in device Settings as a one-time prerequisite. Users should review the action list of any generated shortcut before running it — the same basic check you'd apply to any script.
Viticci puts his own success rate at roughly 90% on generated outputs. He doesn't publish a test methodology, so treat that figure as an informed estimate rather than a benchmark. For bounded tasks — fetch something, process it, send it somewhere — it sounds plausible.
Why natural language Apple Shortcuts usually fail
When Viticci tested drewocarr's January 2026 GitHub skill, an earlier Claude Code approach to the same problem, the outputs regularly included actions that don't exist in Shortcuts, missing variables, and automations that simply failed on import, he reported today. This is the standard failure mode for raw LLM generation against Apple's action catalog: structurally plausible, functionally broken.
Shortcuts Playground's core fix is what Viticci calls the "Craig Loop." A Python validation script runs repeatedly after each file write, feeding structural errors back into the agent's context until the output passes. The plugin also ships a knowledge base built from Viticci's decade of Shortcuts documentation, expanded via Codex from roughly 1,000 actions and intents to roughly 2,000.
ShortcutForge, published three months before today's launch through independent development, reaches the same architectural conclusion. Its documentation explicitly treats LLM output as "untrusted intermediate code" and describes its multi-phase linter, not the model, as "the primary reliability mechanism."
The mechanics are specific: when the model hallucinates an action name, the linter fuzzy-matches it against a 615-action catalog. Unclosed conditional blocks get repaired. Invalid conditions get normalized before the file is ever assembled.
Both projects independently converge on the same answer. The model is the front end; deterministic validation is what makes the output runnable. Without it, an LLM generating structured executable output will produce "almost right" results, and for anything that has to actually run, "almost right" means broken. This pattern generalizes beyond Apple automation to any AI tool operating against a strict schema or execution environment.
How Shortcuts Playground fits into what Apple is already building
Apple announced last year that macOS Tahoe 26 would add a new class of intelligent actions to Shortcuts, including direct access to Apple Intelligence models for tasks like summarizing text and generating images, plus optional ChatGPT integration for broader queries, according to Apple Newsroom. T
he same release added time-based and event-based automation triggers on macOS. A larger action catalog and more trigger types expand the automation surface Shortcuts Playground is already designed to work against.
Viticci says he expects Apple to "Sherlock" this entire category at WWDC in three weeks, and says he wants it to. His stated goal is for Apple to ship an official Apple Intelligence-backed natural-language Shortcuts generator in a future OS release that makes this project unnecessary. The open-source release functions simultaneously as a working tool and a public proof of concept.
Shortcuts Playground arrives after a window of open-source experimentation that also produced drewocarr's earlier skill in January, ShortcutForge's compiler approach in February, and owgit's shortcut-agent-skill in late March. Each project converged on similar architecture. If Apple does absorb this capability natively, it will be inheriting a problem that independent developers have already largely solved, with the technical approach documented publicly.
Whether or not WWDC delivers anything, today's release is the most complete implementation of Claude Code Apple Shortcuts generation available right now. Users who want natural-language automation on Apple's platform before any official version ships have a working option. The architecture argument for why it works treat model output as untrusted, validate deterministically, loop until it passes is the part worth carrying forward regardless of what Apple announces.
Comments
Be the first, drop a comment!