# AI Agent Skills is a directory of skills, and the install verb is the interface > Twenty-four agent skills that work with any client implementing the Agent Skills Standard, installed either as a Claude Code plugin or through the npx skills CLI. What is interesting is the consistency of the design: one directory per skill, read-only defaults on the three database skills, and per-skill OAuth rather than a shared server. **sanjay3290/ai-skills** — 24 cross-platform agent skills for Claude Code, Cursor, Codex & Gemini CLI — databases, messaging, research, TTS, DevOps, and Google Workspace - Repository: https://github.com/sanjay3290/ai-skills - Website: https://deepwiki.com/sanjay3290/ai-skills - Stars: 432 · Forks: 46 - Language: Python - License: Apache-2.0 - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/sanjay3290-ai-skills ## The repository is a directory of skills and almost nothing else Look at the top level and there is no code: .claude-plugin/, .gitignore, LICENSE, README.md and skills/. The product is the directory, and the repository exists to distribute it. That shape follows from the standard it targets. Any client that implements the Agent Skills Standard can read a skills directory, so nothing here has to know which agent is loading it. The cost of that portability is that all the per-client knowledge lives in documentation rather than in adapters. The six clients named with their expected paths are Claude Code, Gemini CLI, Google Antigravity, Cursor, OpenAI Codex CLI and Goose. Each has a user-level directory and a project-level directory, and for three of the six the two are not simply the same path with a tilde in front: Goose keeps project skills in .goose/skills/ and global ones in ~/.config/goose/skills/, and Antigravity uses .agent/skills/ at project level. ## Three database skills that differ only in how they refuse a write Postgres, MySQL and Microsoft SQL Server are three separate skills with three separate defensive descriptions, and the difference between them is the enforcement mechanism rather than the capability. Postgres is described as read-only queries with defense-in-depth security. MySQL is read-only with session-level write protection. MSSQL is read-only with query validation security. So one leans on layered controls, one on session state, and one on validating the statement itself. That difference is not cosmetic. Session-level protection is the weakest of the three in one respect, since anything that opens its own session is outside it, and statement validation is the strongest in a different respect, since it rejects a write regardless of who issues it. If you are choosing between them for a read-only analytics use case, that difference is the decision. All three ship as separate skills rather than one database skill with a dialect flag, which means you install only the engines you actually have. ## Azure DevOps is the outlier at 99 tools across 13 domains Most of these skills describe a small surface. Azure DevOps does not: 99 tools across 13 domains, covering projects, work items, repos, pull requests, pipelines, wikis, test plans, security alerts, variable groups, environments and approvals, branch policies, and attachments. That breadth changes the risk profile of installing it. A skill for Telegram sending alerts and one for reading a wiki are narrow enough that you can reason about what they do. A skill that can modify branch policies, approve environments and change variable groups is closer to an admin credential than a helper, and it will be available to the agent in every session where it is installed. The Atlassian skill is smaller in scope, Jira issues and Confluence pages, and offers a choice of authentication: OAuth 2.1 through an MCP server, or an API token as a fallback. The choice between the two is worth making deliberately rather than accepting the first one that works. ## Google Workspace skills each carry their own OAuth and refuse personal Gmail Seven Google Workspace skills are offered as lightweight alternatives to the full Google Workspace MCP server: Chat, Docs, Sheets, Slides, Drive, Calendar and Gmail. Each one has standalone OAuth authentication with cross-platform token storage through the keyring, which means seven independent credential sets rather than one shared token, and it means revoking one does not disturb the others. It also means you have seven consent screens to walk through rather than one. The stated constraint is hard: a Google Workspace account is required and personal Gmail accounts are not supported. That is because these skills use the same OAuth infrastructure as the official Workspace MCP, and that infrastructure is built around Workspace identities. Anyone planning to use this on a personal account should not expect it to work. Whether this is better than the full MCP server is a matter of footprint: seven skills you install individually are smaller than one server you always run. ## Two text-to-speech skills, two providers, one shared shape ElevenLabs and Google Cloud TTS do the same job at different points on the quality and control curve, and the skill descriptions are more informative than the labels suggest. Both narrate documents and both generate podcasts. ElevenLabs reads PDF, DOCX, MD and TXT and is described as producing two-host podcast generation. Google Cloud TTS is named by its voice families, Neural2, WaveNet and Studio, and by its language count, forty or more. The split that follows from that is document pipeline versus voice library. If your input is a pile of documents and the output is a two-voice audio piece, the ElevenLabs skill is the one described for it. If you want to pick a specific voice and know it is available in your language, the Google skill names the catalogue it draws on. Having both installed is reasonable. They do not conflict, and neither is a general text-to-speech client for interactive use. ## apple-container sidesteps the Docker daemon entirely on Apple silicon The apple-container skill builds, runs and manages OCI and Linux containers as lightweight virtual machines on Apple silicon macOS, using Apple's own container CLI, with no Docker daemon. The distinction matters more than the feature list suggests. A daemon means a background service, a socket and a lifecycle that has to be managed, and on macOS it also historically meant a separate virtualisation stack. Apple's CLI is a different model, and a skill that drives it gives the agent container management without asking you to keep a daemon resident. The other delegation skills follow the same shape of intent. Jules takes async work, bug fixes, documentation, tests and features. Manus takes deep research, market analysis and reports. grok-build runs xAI's Grok Build CLI headlessly, and its description is unusually specific about the division of labour: the assistant plans, writes the specification, dispatches, and reviews every diff. Four of these skills exist to hand work to another agent rather than to do it locally, which is the clearest statement of what a skill can be. ## The install verb is the interface: one skill, many skills, or all of them The npx skills CLI, from vercel-labs, is presented as the package manager for the open Agent Skills ecosystem. Four commands cover the whole surface: ```bash # Browse all 24 available skills npx skills add sanjay3290/ai-skills --list # Install a single skill (auto-detects your agent) npx skills add sanjay3290/ai-skills --skill postgres # Install multiple skills at once npx skills add sanjay3290/ai-skills --skill postgres --skill mysql --skill mssql # Install all skills npx skills add sanjay3290/ai-skills --all ``` The -a flag picks the target, and it repeats: ```bash npx skills add sanjay3290/ai-skills --all -a '*' ``` Left off, -a means auto-detect. With no -g, skills install into the current project directory, which keeps a repository's tooling with that repository. The CLI itself supports more than forty agents, of which seven are tabulated. ## Global versus project scope is one flag, and Claude Code gets a different path entirely Scope is the second decision, and it is a flag rather than a separate command: ```bash # Global install, available in all projects npx skills add sanjay3290/ai-skills --skill imagen -g # Project install (default), scoped to current repo npx skills add sanjay3290/ai-skills --skill imagen ``` Claude Code users have a path that skips the CLI entirely, through its plugin marketplace: ``` /plugin marketplace add sanjay3290/ai-skills /plugin install ai-skills@ai-skills ``` The difference is granularity. All twenty-four skills install as a single plugin, and updates come from /plugin marketplace update ai-skills. So the plugin route is all-or-nothing per repository, while the CLI route is per skill. That distinction decides which one you use. If you want a couple of skills everywhere, the CLI with -g is the right tool. If you want the whole set in one project, the plugin is fewer steps. ## Conclusion Adopt AI Agent Skills if you want a catalogue you can install selectively rather than a platform, because the npx skills CLI takes a single skill, a list of skills, or all of them, and installs into one agent, several, or every supported one. Skip it if you need a Workspace integration on a personal Gmail account, which the Google skills explicitly do not support, and read the three database skills carefully before pointing them at a production instance, since they are documented as read-only but each takes a different approach to enforcing that. Before you commit, confirm the skill count against what you actually need, since the tables here list sixteen general skills and seven Workspace skills while the description and install commands both say twenty-four. Also note the release history: a single tag, v1.2.0 dated 2026-07-09, against a last push on 2026-09-10, so the directory is moving faster than the tags. ## FAQ ### How do you install Claude Code AI skills from this collection? Through the plugin marketplace with /plugin marketplace add sanjay3290/ai-skills followed by /plugin install ai-skills@ai-skills, which installs all the skills as a single plugin and updates later with /plugin marketplace update ai-skills. For finer control, the npx skills CLI installs individual skills instead. ### Which AI clients can use these agent skills? Any client implementing the Agent Skills Standard. Six are named with their directories: Claude Code at ~/.claude/skills/ or .claude/skills/, Gemini CLI, Google Antigravity, Cursor, OpenAI Codex CLI and Goose. The npx skills CLI itself supports more than forty agents. ### Are the database skills in AI Agent Skills read-only? They are described as read-only, and each enforces that differently: Postgres with defense-in-depth security, MySQL with session-level write protection, and Microsoft SQL Server with query validation. They ship as three separate skills rather than one skill with a dialect option. ### Can I use the Google Workspace skills with a personal Gmail account? No. They are described as lightweight alternatives to the full Google Workspace MCP server, each with standalone OAuth authentication and cross-platform token storage via keyring, and they require a Google Workspace account. Personal Gmail accounts are explicitly not supported. ### How many skills does the sanjay3290 ai-skills collection contain? The repository description and the install commands both say twenty-four. The tables visible in the documentation list sixteen general skills covering databases, image generation, research, wikis, delegated agents, text-to-speech, Atlassian, Azure DevOps, containers and Telegram, plus seven Google Workspace skills. ## Sources - [License: Apache-2.0](https://github.com/sanjay3290/ai-skills/blob/main/LICENSE) - [Project website](https://deepwiki.com/sanjay3290/ai-skills) - [README](https://github.com/sanjay3290/ai-skills/blob/main/README.md) - [Releases](https://github.com/sanjay3290/ai-skills/releases) - [sanjay3290/ai-skills on GitHub](https://github.com/sanjay3290/ai-skills) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/sanjay3290-ai-skills