Agent Skills for driving Mac apps and dev workflows from an AI agent. Works with any tool that supports the Agent Skills spec, including Claude Code.
npx skills add ljagiello/agent-skillsWant these skills as part of a real workflow — schedules, signals, MCP tools, memory, the works? Drop them into Friday, the shareable AI workspace runtime from Tempest Labs.
Friday Studio loads skills into agent context on demand and runs them inside reproducible workspaces that you can trigger from chat, on a cron, or over HTTP. Everything runs locally, your data stays on your machine, and every step is logged so you can see exactly what the agent did.
To add these skills to Friday Studio:
- Install Friday from hellofriday.ai (macOS).
- Open Skills in the Studio sidebar and click + Add.
- Import individual skills by reference (e.g.
ljagiello/agent-skills/utmapp), or upload this repo as a folder. - Reference them from any
workspace.yml, or let agents load them automatically based on the skill description.
See the Friday Skills docs for the full workflow, and the Friday blog — including AI Drift: The Hidden Cost of Building with AI — for the philosophy behind it.
| Skill | Files | Description |
|---|---|---|
| utmapp | 6 | Drive UTM virtual machines on macOS via utmctl and AppleScript / JXA — list/create/start/stop/suspend/clone/delete VMs, run commands inside guests, transfer files, query guest IPs, send keyboard or mouse input, forward USB devices, hand-edit .utm config.plist, and walk through installing Linux / Windows 11 ARM / Windows on Intel / macOS guests. Covers both the QEMU backend (cross-architecture emulation) and the Apple Virtualization backend. |
| cleanshotx | 2 | Drive CleanShot X on macOS via its cleanshot:// URL scheme — capture area / window / fullscreen / scrolling / previous-area screenshots, record the screen, OCR text from images or screen regions, pin images, open files in the annotator, manage history and Quick Access overlays, toggle desktop icons, and open specific settings tabs. Covers all 19 commands with full parameter and version-gating reference. |
Skills are loaded automatically by the agent based on the prompt — mention UTM, utmctl, or a .utm bundle and the utmapp skill activates; mention CleanShot, cleanshot://, scripted macOS screenshots, or screen-region OCR and cleanshotx activates. The agent reads each skill's SKILL.md first and then pulls in its domain-specific reference files only as needed for the task at hand.
Each skill follows the Agent Skills specification: a SKILL.md with YAML frontmatter at the top of the skill directory, plus optional references/, scripts/, and assets/ subdirectories. Reference files are loaded on demand so the metadata budget stays small.
When adding or editing a skill, every factual claim should trace back to a verified source — for utmapp that meant UTMCtl.swift, UTM.sdef, and the Configuration/*.swift CodingKeys declarations in the upstream UTM repo. Treat the spec's authoring best practices as the bar.
Two layers of automated checks live under evals/:
- Static validator (
python3 evals/validate.py) — stdlib-only, deterministic, free. Verifies frontmatter limits, name/dirname match, body line count, and that every Markdown reference resolves to a real file. - Behavioral evals (
npx promptfoo eval --config evals/promptfooconfig.yaml) — drive Claude with each skill loaded as the system prompt and assert that it emits valid commands, applies the skill's gotchas, and refuses to invent missing features. RequiresANTHROPIC_API_KEY.
See evals/README.md for the full setup, how the prompt function wires SKILL.md into the system prompt, and how to add cases for a new skill.
MIT