Azure Agent Skills (Microsoft Official Skill Library)
Curated, pre-compiled Azure expertise from Microsoft Learn docs that AI coding assistants like Claude Code, GitHub Copilot, and Cursor load on demand.
Least privilege: the repository ships Markdown skills and plugin manifests; no executable scripts or destructive operations are shown. One point deducted because skills declare network access to fetch Microsoft Learn docs with no defined fetch scope. User confirmation: no guidance anywhere about asking the user before performing actions. Data flow transparency: clearly documented that skills fetch Learn content via MCP (mcp_microsoftdocs) or fetch_webpage; flow is clear but caching/logging behavior is unstated. Sensitive data handling: entirely unaddressed — no guidance on secrets, connection strings, or prompt-injection risk from fetched pages. Dependency security: no executable dependencies; relies on the host assistant's fetch/MCP tooling — low risk but no supply-chain statement. External effects: skills are documentation guidance with no direct side effects. Rollback: manual install is deletable, plugins uninstallable; README even suggests deleting the clone — adequate but no version pinning or downgrade path. Source attribution: every skill points to official Microsoft Learn links; attribution handled well, full marks.
Self-consistency: README tables (skill counts, paths, manifest locations) are internally consistent, but consistency of the 193 skill files themselves cannot be checked from this evidence. Dependency availability: heavily dependent on the external Learn MCP server and network fetch; degraded behavior is undefined. Failure messages: no documented strategy for failed fetches, dead links, or unavailable MCP tooling — clear deduction.
Audience and scenarios: role-based bundles and a multi-assistant path table show strong audience thinking, but actual per-scenario coverage inside skill files is unverifiable here. Capability boundaries: the repo does not state what skills cannot do (e.g., no Azure CLI execution, no resource modification). Trigger precision: triggering relies on YAML frontmatter descriptions, but no SKILL.md was included, so description quality cannot be verified. Environment fit: paths and enablement for 8 assistants, including the VS Code experimental chat.agent.skills toggle, are documented in detail — near full marks.
Information architecture: categorization by Azure service domain with a clear path/responsibility table (skills/, products/, plugin manifests). Install notes: exemplary — one-click plugin paths, manual copy with correct/incorrect examples, and the experimental VS Code flag. Naming stability: naming conventions described, but weekly crawls may add/remove/rename skills with no stability commitment. Examples and FAQ: usage examples and path caveats exist; no formal FAQ. Known limitations: VS Code preview status, experimental flag, and contribution policy restrictions are flagged. License: dual CC-BY-4.0 (docs) + MIT (code), full license text included. Versioning/changelog: no CHANGELOG or per-skill versions; plugin update mechanics are vague — deducted. Maintenance responsibility: weekly automatic crawl, no manual skill contributions, MSRC reporting channel — clear paths, though no named maintainers or SLA.
Output usability: skill output is defined as doc links + best practices + architecture patterns — a clear structure, but actual per-skill output quality cannot be statically verified. Marginal value: pre-compiling Learn docs into structured skills to avoid repeated RAG is a genuine value proposition, but the content is fundamentally a reorganization of public documentation — one point deducted. Cost benefit: progressive disclosure controls token cost, yet fetching full doc pages per query has unstated token overhead.
Claim traceability: skills claim Learn provenance with links — good design, though links are not individually checked in this evidence. Cross-source corroboration: references to official Learn/Well-Architected sources give moderate corroboration potential; content itself not included. Fact/inference separation: no SKILL.md was provided in this evidence set, so whether skills separate documented facts from author inference cannot be verified — this is the largest gap in the review.
- All skills depend on the Microsoft Learn MCP server or a web fetch tool; skills degrade to uselessness when these are unavailable, and no fallback behavior is defined.
- Fetched external page content flows directly into the assistant's context, creating prompt-injection risk; the repo provides no sensitive-data handling or injection guidance.
- Publisher identity is not verified by the FollowAgents enterprise registry; confirm provenance and content before enterprise deployment.
- Skills are generated by weekly automated crawls and may change or disappear without notice; there is no version pinning or changelog — unsuitable for environments requiring a stable skill set.
- This is a static source review with no execution; per-SKILL.md content quality and fact/inference separation could not be verified.
What does this agent do, and when should you use it?
This open-source repository, maintained by the MicrosoftDocs GitHub organization, ships 193 skills following the Agent Skills open standard (SKILL.md format) across 19 Azure service categories, including Compute, Data & Analytics, AI/ML, Networking, Security, and Infrastructure. Skills are not standalone programs; they are Markdown instructions and resources that AI coding assistants load on demand via three-level progressive disclosure (discovery, instructions, resources). The repo also carries native plugin manifests for VS Code, Claude Code, and OpenAI Codex, enabling one-click installation of all skills. Most skills require network access and fetch up-to-date official documentation through the Microsoft Learn MCP Server or a web fetch tool at usage time. Documentation content is CC BY 4.0 licensed and code is MIT licensed; manual skill contributions are not accepted — content updates flow from Microsoft Learn documentation via an automatic weekly crawl.
The repository pre-compiles decisions, procedures, best practices, and constraints from Microsoft Learn documentation into structured skills: each skill is a folder containing a SKILL.md with YAML frontmatter (name, description) plus curated doc links, architecture patterns, and troubleshooting guidance. Once installed, an assistant discovers skills from frontmatter, loads the full SKILL.md when triggered, and accesses scripts and examples as needed. At runtime the assistant fetches live documentation via the mcp_microsoftdocs:microsoft_docs_fetch tool (Microsoft Learn MCP Server) or a fetch_webpage tool, enabling answers to questions like how to implement blue-green deployments in Azure Container Apps. The skills/ directory holds production-ready skills; plugin., .claude-plugin/, and .codex-plugin/ provide per-platform plugin manifests; products/ stores internal scan-pipeline data.
- An Azure developer using Claude Code or GitHub Copilot wants the assistant to automatically know official best practices for services like Functions and Container Apps without pasting docs into every conversation
- An AI/ML engineer installs the AI/ML Developer bundle to quickly load skills for Azure AI Services, Microsoft Foundry, and related services
- An infrastructure engineer uses the Infrastructure Pro bundle for guidance on networking, backup, recovery, and resource management
- A security and compliance team loads the Security & Compliance bundle covering RBAC, policy, key management, and compliance
- A team standardizes skill delivery via the plugin model, configuring chat.plugins.marketplaces in VS Code to receive weekly refreshes crawled from Learn docs
- Users of Cursor, Gemini CLI, OpenCode, or AdaL CLI copy the skills/ contents to their tool's path (e.g., .cursor/skills/) and invoke skills via natural language or @skill-name
What are this agent's strengths and limitations?
- Content is pre-compiled from official Microsoft Learn documentation with an automatic weekly crawl, keeping guidance aligned with current best practices
- One repository ships native plugin manifests for VS Code, Claude Code, and OpenAI Codex — a single install delivers all 193+ skills with built-in updates
- Works with Claude Code, Gemini CLI, Codex CLI, Antigravity, GitHub Copilot, Cursor, OpenCode, and AdaL CLI via the agentskills.io open standard
- Three-level progressive disclosure loads only what is relevant, avoiding context bloat
- Ten role-based bundles (Quick Start, Popular, AI/ML Developer, etc.) lower onboarding cost
- Most skills require network access to Microsoft Learn (MCP Server or web fetch), making them largely unusable offline
- No manual skill creation or editing is accepted; content changes require updating Learn documentation plus a re-crawl request or waiting for the weekly crawl
- VS Code agent plugins and Agent Skills are preview/experimental features requiring manual flags (chat.plugins.enabled, chat.agent.skills), possibly blocked by organization policy
- Skills are documentation-style instructions, not executable code; output quality depends on the assistant's interpretation and live fetching
- Coverage is exclusive to the Microsoft/Azure ecosystem; manual installs have no auto-update and must be re-copied to refresh
How do you install or deploy this agent?
Recommended plugin install:
- VS Code: Command Palette → "Chat: Install Plugin From Source" → paste https://github.com/MicrosoftDocs/agent-skills; or add "MicrosoftDocs/agent-skills" to chat.plugins.marketplaces in settings., then install azure-agent-skills from the Extensions view via @agentPlugins (requires the chat.plugins.enabled preview flag).
- Claude Code: run /plugin marketplace add MicrosoftDocs/agent-skills, then /plugin install azure-agent-skills@microsoft-agent-skills, then /reload-plugins.
- Codex: run codex plugin marketplace add MicrosoftDocs/agent-skills, then install azure-agent-skills from /plugins.
Manual install: git clone https://github.com/MicrosoftDocs/agent-skills.git, then copy the folders inside skills/ (e.g., azure-functions/) to your assistant's path — Copilot: .github/skills/ or ~/.copilot/skills/; Claude Code: .claude/skills/ or ~/.claude/skills/; Codex: .codex/skills/; Cursor: .cursor/skills/; Gemini CLI: .gemini/skills/; OpenCode: .agent/skills/. For VS Code manual mode, enable "Chat: Use Agent Skills" (chat.agent.skills) in Settings. Copy the folders inside skills/, not the skills folder itself.
How do you use this agent?
Ensure your assistant has network access (most skills need it to fetch current Microsoft Learn docs; the Microsoft Learn MCP Server's mcp_microsoftdocs:microsoft_docs_fetch is recommended, or a fetch_webpage tool). Then just ask naturally: "Help me set up Azure Functions with Durable Functions", "What are the best practices for Azure Container Apps?", "How do I configure Azure API Management?". The assistant discovers relevant skills, loads instructions via progressive disclosure, and fetches live documentation. Invocation styles: natural language or /skill-name in Claude Code, @skill-name in Cursor, opencode run @skill-name in OpenCode, @workspace in Copilot. Consult docs/BUNDLES.md for role-based bundles like Quick Start and Popular.
How does this agent compare with similar options?
The README lists it alongside anthropics/skills (Anthropic's official document-handling skills), openai/skills (the Codex skills catalog), and github/awesome-copilot (a community collection for Copilot); unlike those general-purpose skill repos, this one is dedicated to Azure cloud domain expertise.