claude-skills: A 388-Skill Library for 13 Coding Agents, With One Clear Trade-Off

What This Repository Actually Solves

Coding agents like Claude Code, Codex, and Gemini CLI come with generic capabilities. They can write code, but they do not know your company's compliance workflow, your SEO checklist, or how to run a clinical research budget. claude-skills addresses that gap by packaging domain expertise into SKILL.md files, Python scripts, and reference documents. The README states it contains 388 skills across 20 domains, plus 30+ agents and 70+ custom commands. The intended user is someone who wants to give their agent a specific competency, such as "run a security audit" or "think like a startup CTO," without writing the instructions from scratch. It is a library of reusable expertise, not a single tool. For an engineer, the immediate appeal is the breadth: one repository covers engineering, marketing, compliance, finance, and C-level advisory. The trade-off is that breadth often comes at the cost of depth per skill.

How the Skills Are Structured and Why That Matters

Each skill follows the agentskills.io SKILL.md standard. The README explains that a skill includes a SKILL.md file with structured instructions, workflows, and decision frameworks, plus Python tools and reference docs. The Python tools are all stdlib-only, meaning no pip installs are required. That is a significant design choice. It lowers the barrier to running the scripts on any machine with Python, but it also limits what the tools can do. No third-party libraries means no advanced data processing, no machine learning, no API wrappers. The skills are essentially instruction packages plus simple scripts. The repository also distinguishes between skills, agents, and personas. Skills define how to execute a task, agents define what task to do, and personas define who is thinking. This separation is useful for orchestration, as you can combine a persona with a skill to get both voice and procedure. However, the README does not provide a concrete example of that orchestration, so you are left to infer how to compose them.

Installation Across 13 Tools: The Good and the Confusing

Installation varies by tool. For Claude Code, you add a marketplace and install plugins by domain, for example: /plugin marketplace add alirezarezvani/claude-skills, then /plugin install engineering-skills@claude-code-skills. For Gemini CLI, you clone and run ./scripts/gemini-install.sh, then call activate_skill(name="senior-architect"). For Codex, you use npx agent-skills-cli add alirezarezvani/claude-skills --agent codex. Other tools like Cursor, Aider, and Windsurf use a conversion script: ./scripts/install.sh --tool cursor --target . . The README claims conversion takes about 15 seconds for all skills. That speed is plausible because the conversion is mostly format translation, not content generation. The confusing part is the number of tools and formats. The table lists 13 tools, but some are BYO-sync tier, meaning you must run a sync script manually. Hermes Agent and Mistral Vibe ship pre-generated trees but require local installation. This fragmentation means you must read the installation notes carefully for your specific tool, especially on Windows, where symlinks are required for mirror trees.

The Conversion Script: A Real Mechanism, But Check the Output

The multi-tool support relies on ./scripts/convert.sh --tool all, which generates native formats for each agent. For Cursor, it produces .mdc rules; for Aider, CONVENTIONS.md; for Kilo Code, .kilocode/rules/. The README gives a verification step: find .cursor/rules -name "*.mdc" | wc -l should show 346. That number is specific and testable. The conversion is a one-way translation from SKILL.md to each tool's format. The risk is that not all SKILL.md content translates cleanly. For example, Aider's CONVENTIONS.md is a flat file, while a skill may contain multiple sections and scripts. The README does not specify how scripts and references are handled in each target format. It says "Support for scripts, references, templates where applicable," which is vague. If you rely on a skill's Python tool, you need to verify that the converted format actually includes the script and that the agent can execute it. The --force flag skips confirmation, which is convenient for automation but also means you might overwrite existing rules without review.

What the Repository Does Not Tell You: Depth and Maintenance

The README boasts 5,200+ GitHub stars and calls itself the most comprehensive library, but it does not provide evidence of skill quality. Star counts are not a proxy for correctness. The description lists 345 skills, but the README says 388, and the release notes for v2.12.0 mention 380 skills across 20 domains. These numbers do not align. That inconsistency is a red flag for a repository that is updated frequently, with the last push on 2026-08-25. Frequent releases mean the skill set is changing, which is good for freshness but bad for stability. If you install a skill and the next release changes its behavior, your agent's workflow may break. The README also does not document how to update skills after installation, except for the conversion script. There is no mention of version pinning or rollback. For a library this large, you should expect some skills to be outdated or incomplete, especially in fast-moving domains like AI/ML or security. The repository is a catalog, not a curated set of guarantees.

A Real Alternative: Write Your Own SKILL.md Files

The most direct alternative is to create your own skills. The agentskills.io standard is public, and you can write a SKILL.md file for your specific workflow in an afternoon. The advantage is precision: your skill will match your exact toolchain, your exact compliance rules, and your exact coding style. claude-skills gives you generic expertise, but generic expertise may not fit your context. For example, a marketing skill from the repository might include an SEO checklist, but your product's SEO requirements could differ. Writing your own skill also avoids the conversion overhead and the risk of format translation errors. The cost is time and maintenance. You must keep your skill updated as your processes change. The repository's value is that it saves you that effort for common domains. If you only need a few skills, writing them yourself is often simpler than installing a 388-skill library and managing its updates.

Licence and Maintenance Cost: What You Are Signing Up For

The repository is MIT licensed, which means you can use, modify, and redistribute the skills freely, including in commercial projects. That is a low-risk licence for adoption. The Python tools are stdlib-only, so there are no dependency conflicts. The maintenance cost is twofold. First, the repository is large and changes frequently, with releases like v2.9.0 and v2.12.0 coming months apart. You will need to track updates if you want fixes and new skills. Second, the conversion scripts and installation paths are specific to each tool, and they may break when a tool updates its skill format. The README does not mention a changelog or migration guide for major releases. The release notes for v2.12.0 mention a "full issue-triage sweep," which suggests the maintainer is active, but it also implies there were many issues to triage. Before committing, you should check the git history for the specific skills you plan to use to see how often they are updated and whether they have open issues.

Editorial conclusion

Adopt claude-skills if you work across multiple coding agents and want a single source of domain expertise without writing your own SKILL.md files. The MIT license and stdlib-only Python tools keep integration low-risk. Do not adopt it if you need deep, battle-tested workflows in one narrow domain, or if you cannot tolerate a large, frequently updated repository where individual skills may be thin. Before adopting, verify that the skills you actually plan to use are maintained by checking their last commit dates, and confirm that the conversion scripts produce correct formats for your target tool, especially for Cursor and Aider, where the generated rules may not preserve all skill logic. The repository is a broad catalog, not a curated set of guarantees.

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes