Google Workspace CLI Skills
The structured skill catalog that ships with the Google Workspace CLI — 19 core API services, 25 helper shortcuts, 10 role-based personas, and 40 multi-step recipes. This isn't just an API wrapper; it's a designed taxonomy for making Google Workspace legible to AI agents.
Key Quotes
"Auto-generated by
gws generate-skills. Do not edit manually."
The skills are themselves generated from structured source files. This means the skill definitions are the source of truth, and the catalog is a build artifact. The inversion is subtle but important: they didn't hand-curate a skills page — they built a compiler that produces it. #compound-engineering
"One CLI for all of Google Workspace — built for humans and AI agents."
From the parent project. The skills catalog is the agent-facing half of this promise. Humans get gws drive list; agents get gws-drive as a skill with structured instructions for how to use it.
Skill Architecture
The catalog reveals a four-layer design, each building on the layer below:
Services (19 skills)
Low-level API wrappers. Each maps to a Google Workspace product: Drive, Gmail, Calendar, Sheets, Docs, Slides, Tasks, People, Chat, Classroom, Forms, Keep, Meet, Events, Model Armor, Workflow, Script, Admin Reports, plus a shared cross-cutting skill.
Helpers (25 skills)
Single-purpose convenience commands that reduce API friction. Instead of constructing a gws gmail send invocation from scratch, an agent uses gws-gmail-send which bakes in sensible defaults. Notable patterns: NDJSON streaming for real-time event watching (gws-gmail-watch, gws-events-subscribe), and Model Armor safety filtering for both prompts and responses.
Recipes (40 skills)
Multi-step workflows expressed as named skills. These cross service boundaries: recipe-post-mortem-setup touches Docs, Calendar, and Chat. recipe-create-events-from-sheet reads Sheets rows and creates Calendar entries. Each recipe is a codified process — the kind of thing a human writes as a checklist, encoded so an agent can execute it deterministically.
Personas (10 skills)
Role-based bundles that compose services, helpers, and recipes into coherent job functions. An exec-assistant persona gets calendar management + email triage + meeting prep. A researcher gets document organization + references + collaboration. Personas are pre-composed agent configurations — effectively "hire this agent for this role."
Key Themes
The taxonomy itself is the design artifact. Services → Helpers → Recipes → Personas is a hierarchy of increasing abstraction that mirrors how humans think about work. You don't "use the Drive API" — you "share a folder with the team." Recipes encode that intent-to-API mapping.
The auto-generation note (gws generate-skills) implies a meta-tool architecture: the skill files are the source, the catalog is compiled output, and the CLI reads them at runtime. This is the same pattern as the CLI's dynamic Discovery Service integration, applied one level up to the agent interface.
Critical Analysis
The persona concept is the most interesting and least discussed part of this. Everyone builds API wrappers. Few people think about composing them into roles. The 10 personas cover common knowledge-worker archetypes — exec assistant, project manager, HR, sales, IT, content, support, events, team lead, researcher. These aren't arbitrary; they're the organizational chart expressed as tool configurations.
The missing persona is "engineer." No dev-focused role, no CI/CD workflows, no repository management. This is a deliberate boundary: Google Workspace is the office productivity surface, not the engineering surface. But it leaves open the question of what happens when you want a persona that bridges both — e.g., "oncall engineer" who needs Calendar for scheduling, Chat for incident comms, AND GitHub for code.
The recipe count (40) versus persona count (10) is telling. Recipes are cheap to add — each is a linear script. Personas require curation and judgment about which skills belong together. The ratio suggests personas are intentionally constrained while recipes grow unbounded. Smart: personas are the API contract, recipes are the long tail.
Compared to MCP-based Google integrations, this CLI skills approach is simpler to deploy (one binary, shell invocation) but stateless (each call is independent). Recipes bridge this gap by encoding multi-step state machines, but they're static sequences — they can't branch dynamically based on intermediate results the way an MCP server with tool chaining can. For deterministic workflows, that's fine. For exploratory agent behavior, it's limiting.
The Model Armor integration (3 skills: sanitize-prompt, sanitize-response, create-template) is a quiet but significant choice. Including safety filtering at the skill level means agents using these skills get guardrails by default, not as an afterthought. Most agent toolkits don't even have a concept of "sanitize this before sending to the API." #guardrails
Sources: Google Workspace CLI Skills, Google Workspace CLI Last updated: 2026-05-15