The business process automation tool stopped being a .xaml file, and nobody wrote the memo

The textbook answer is software that moves a repeatable business process off a human's plate, usually by stitching together clicks across apps. That definition was written for the recorder era, when the tool was a .xaml workflow in UiPath or a ScriptRunner macro in Automation Anywhere. The 2026 answer is different on an LLM-driven desktop agent. The tool is whatever the model reads before it acts. For Fazm on macOS, that is a markdown file with a YAML frontmatter header and a few hundred lines of plain English. No canvas, no nodes, no recorded trace.

It does when the runtime is an LLM. The LLM reads the skill file into its context window, sees a description plus a quick-start table plus the shell commands and SQL snippets the author included, and then drives the user's Mac through the macOS accessibility tree. The skill file plays the same role a .xaml plays in UiPath: it is the unit of automation you edit, check into git, and ship. But because the runtime is reading prose rather than executing a flow graph, the author writes English instead of wiring nodes. That collapses the build step into writing a runbook.

Seventeen. They live at Desktop/Sources/BundledSkills/ in the Fazm desktop repo and get copied into the signed app bundle at build time. The list today: ai-browser-profile, canvas-design, deep-research, doc-coauthoring, docx, find-skills, frontend-design, google-workspace-setup, pdf, pptx, social-autoposter, social-autoposter-setup, telegram, travel-planner, video-edit, web-scraping, xlsx. Each is a single *.skill.md file. There is no separate binary for each one. The LLM reads them on demand.

On launch, Fazm runs SkillInstaller.swift. The installer scans Bundle.resourceBundle/BundledSkills/ for files whose path extension is "md" and whose filename also ends in ".skill", builds a list of skill names, and for each one reads a SHA-256 hex digest of the bundled file using CryptoKit's SHA256.hash. It compares that digest against the file currently at ~/.claude/skills/<name>/SKILL.md. Same digest, nothing happens. Different digest, the old file is removed and the bundled file is copied in. That is it. The whole update pipeline for all 17 skills is about 170 lines of Swift.

Open Desktop/Sources/SkillInstaller.swift in the Fazm repo. Line 30 to 40 is the auto-discovery: it filters `contentsOfDirectory` for `pathExtension == "md"` then for `deletingPathExtension().pathExtension == "skill"`. Line 58 to 62 is the digest function: `let digest = SHA256.hash(data: data); return digest.map { String(format: "%02x", $0) }.joined()`. Line 100 to 141 is the install/update loop that compares the two digests and either skips, overwrites, or installs. Line 177 to 178 is the small npm-skill registry for skills that ship outside the bundle (today: social-autoposter).

Those tools store the automation unit in a binary or semi-binary artifact that only their own designer can edit. UiPath workflows are XAML. Automation Anywhere bots are .atmx. Zapier zaps are proprietary JSON. Make scenarios are a blueprint JSON. Power Automate flows are a cloud-stored JSON graph. All five require their own editor to author and their own runtime to execute. A Fazm skill is a plain text file that anyone can read in any editor and that runs inside any LLM that loads the file. The author is the LLM user; the runtime is the LLM.

Yes, and that is the design. A skill file starts with a YAML frontmatter header (`name`, `description`, optional `user_invocable`), followed by a short markdown body that tells the model how to behave. The bundled skills range from 42 lines (frontend-design) to 856 lines (deep-research). The pattern you copy is: describe the intent, list the steps the model should take, include example shell commands or SQL, note the apps it will drive through the accessibility tree. That is also how a junior teammate would write a runbook for a senior one. The skill file and the runbook converge.

Nothing replaces it because nothing is recorded. You describe what you want done. The model reads the skill, queries the live macOS accessibility tree with AXUIElementCreateApplication and kAXChildrenAttribute, and acts on the real structural elements (AXButton, AXTextField, AXMenuItem) rather than the remembered pixel locations of a previous run. When Apple reskins Mail in the next macOS point release, a recorder-era automation breaks. A skill-plus-tree automation keeps working because the roles and labels have not moved, only the chrome.

Your edits live at ~/.claude/skills/<name>/SKILL.md. The next time Fazm launches, if the bundled version inside the app has not changed since the last install, your edits survive: the installer's SHA-256 check of the bundled file against its own prior state is a no-op path. If the bundled version has changed (you installed a new Fazm release), your edits are overwritten with the new bundled skill, and a toast fires telling you which skills were updated. This is intentional: skills-in-bundle are treated as first-party code that the app owns. Fork the skill outside the bundled-name list if you want total control.

The Fazm desktop app is public. The 17 .skill.md files are in Desktop/Sources/BundledSkills/. The installer that picks them up is Desktop/Sources/SkillInstaller.swift. The MCP server it pairs with (the binary that reads the macOS accessibility tree at runtime) is mediar-ai/mcp-server-macos-use. The agent communication bridge that wires LLMs to both is in acp-bridge/src/index.ts in the same repo. Every claim in this guide is a `grep` away.