Deployment Options

The question this page answers: "I have a skill I want to use (or roll out). Where does it go?"

Two questions drive the decision:

1. Who is using it? Just you, a project team, or the whole company?
2. Whose environment? An APD account/machine, or a client's?

Decision tree

Pills: Self-serve = minutes, you own it · Via repo = team-wide through the project repo · Coordinated = company-wide, needs IT / admin / marketplace coordination

The deployment mechanisms in detail

1. Local folder (self-serve)

Copy skill folders into the agent's skills directory. Fastest path; you own updates (git pull + re-copy).

APD-managed Windows machines: Claude Code runs sandboxed in WSL, and the Windows-profile C:\Users\[MYUSER]\.claude is not visible. Use the WSL side instead: \\wsl.localhost\Ubuntu\home\[MYUSER]\.claude for all projects, or the .claude/skills/ of the folder WSL is active in. Setup and full restrictions: Claude Code on Windows. Codex users can instead run the built-in $skill-installer from a session in this repo; it prompts for the target path (see the repo README quick start).

Agent	Project-level	User-level (all projects)	Invoke
Claude Code	.claude/skills/	~/.claude/skills/	/skill-name
Codex	.agents/skills/	~/.agents/skills/	$skill-name
Cortex Code	.cortex/skills/ or .claude/skills/	~/.snowflake/cortex/skills/ or ~/.claude/skills/	/skill-name
Genie Code (Databricks)	—	/Users/{username}/.assistant/skills/ (Unity Catalog)	@skill-name

Both Claude Code and Codex discover skills hierarchically from the current directory up to the repo root, so a top-level folder covers everything beneath it. Full provider comparison: ADR-002.

Team variant: commit the project-level folder into the project repo. Everyone who clones gets the skills: no per-person setup, changes go through PR review, and removal is a delete. This is the recommended default for project teams, on APD and client repos (client policy permitting).

2. Plugin for Claude Code

Package skills as a Claude Code plugin distributed through a plugin marketplace (a git repo with a .claude-plugin/marketplace.json). Users install with:

/plugin marketplace add <org>/<marketplace-repo>
/plugin install <plugin-name>@<marketplace-name>

A plugin can bundle multiple skills (plus commands and other components) and is much easier to version and update than hand-copied folders; this is how Anthropic distributes its own document skills.

Best for: rolling out a curated skill set to many Claude Code users without touching each machine.

3. Enterprise / chat-product upload

Getting skills into the chat products (not the CLI agents):

• claude.ai (individual / client accounts): custom skills are uploaded per user in Settings › Capabilities. Currently blocked on APD-managed machines (see Claude Code on Windows).
• Claude Enterprise: org-level skills managed in the workspace and pre-loaded for every chat user; this is how apd-pptx-design ships internally.
• ChatGPT chat: no per-user upload; skills reach chat users via the Enterprise workspace rollout, coordinated with the workspace admin per OpenAI's current enterprise controls.

Chat-product runtimes also differ from CLI agents (e.g. restricted network and package installation), so test a skill in the target surface before recommending it there.

Best for: making skills available to APDers who work primarily in the chat UI rather than a terminal.

4. File push by MSFT admin (Intune)

IT pushes skill folders to a standard path (e.g. ~/.claude/skills/) on managed APD laptops. Zero user effort; updates ship on IT's schedule.

Best for: guaranteeing a baseline skill set (e.g. brand voice, deck design) on managed APD machines, including non-developers; unmanaged users self-serve from the repo. Raise via IT / the AI Knowledge Hub Support Desk.

Quick chooser

Scenario	Path
"I want to try apd-pptx-design today"	🟢 Local user-level folder
"I mostly use claude.ai, not the terminal"	🟢 Upload to claude.ai Settings › Capabilities
"My delivery team should all use our modeling skill"	🔵 Commit .claude/skills/ (or .agents/skills/) into the project repo
"We're on a client's tenant"	🔵 Policy + SoW/MSA check → main-committed project folder
"Roll out a baseline to groups in Aimpoint"	🟠 Intune push (managed laptops) + 🟢 self-serve for the rest
"Distribute a skill set to all Claude Code users"	🟠 Plugin marketplace
"Claude chat users company-wide"	🟠 Claude Enterprise org-level skills
"Codex users company-wide"	🟠 Enterprise workspace rollout with the admin

Client account vs APD account

Before deploying anything into a client environment:

• Confirm the client's AI usage policy and the SoW/MSA language allow agent skills, and check what data the skill touches
• Deploy only skills committed to main, at project level: visible, reviewed in PRs, removable
• Skills marked license: Proprietary need approval before they go into a client tenant
• When in doubt: Office Hours or the Support Desk