A free, open-source library called claude-skills has grown into the most comprehensive collection of reusable skill packages for AI coding agents, shipping more than 345 production-ready packages that work without modification across 13 tools including Claude Code, OpenAI Codex, Gemini CLI, Cursor, Hermes Agent, and Windsurf. The library, maintained by developer Alireza Rezvani and updated to v2.10.3 on June 10, draws attention because it solves a problem that now affects most engineering organizations: when developers on the same team use different AI coding agents, there is no standard way to make all those agents behave like domain experts. The library provides one.
The deeper significance goes beyond convenience. As AI coding agents spread across the industry and teams adopt multiple tools in parallel, whoever defines the standard context engineering packages for engineering domains becomes the de facto definer of what “senior engineer behavior” means in AI-assisted development workflows — regardless of which underlying model the agent runs on.
What an AI Coding Agent Skill Actually Does
The SKILL.md open specification, which Anthropic published in December 2025, is the format all 13 supported tools share. A skill is not a prompt. It is a folder containing a SKILL.md instruction file, a references/ directory with detailed domain guides, and in many cases a scripts/ directory containing Python automation tools the agent can invoke directly.
The loading mechanism is architecturally important. At the start of each session, an agent scans every available skill’s name and description — these take only a few dozen tokens each — and loads the full content of a skill only when the user’s request triggers it. That progressive disclosure design means a team can install hundreds of skills without flooding the model’s context window on every interaction. The key is that the description field in each SKILL.md is written for the model, not for human readers: it contains the specific trigger phrases and contexts that signal the skill is relevant.
Every Python script bundled with a claude-skills package runs on the standard library alone: no pip installs, no external API keys, no LLM calls. That constraint is deliberate. A skill that requires external services is fragile; a skill built entirely from Python’s standard library runs anywhere Python runs and introduces no supply chain dependencies.
How Cross-Agent Compatibility Works
The library’s portability is not the result of a single universal format. Claude Code, Codex, and Gemini CLI each read skill files from different directory paths and in slightly different configurations. The claude-skills repository ships conversion scripts — scripts/convert.sh for individual skills and per-agent install scripts such as scripts/gemini-install.sh — that translate the Claude Code-native layout into each agent’s expected structure. A scripts/sync-vibe-skills.py script handles Mistral Vibe; for Codex, the scripts create .codex/skills/ symlinks. For Claude Code users, the catalog is accessible through the native plugin marketplace.
The result is that a team can define one authoritative set of engineering behaviors and propagate them across every coding agent in use. A 2026 DX Report survey of engineering leaders found that 73% cited inconsistent behavior across their team’s AI tools as a primary productivity problem; this library’s cross-agent installation pattern directly addresses that gap.
What Are the 51 Engineering-Team Skills?
The library’s engineering-team bundle has grown to 51 core skills, covering the full span of a modern software organization. Roles include senior architect, frontend, backend, fullstack, QA, DevOps, SecOps, data engineer, ML engineer, prompt engineer, and computer vision specialist. Each package follows a consistent layout: a SKILL.md instruction file, a references/ folder containing detailed guides on domain patterns and best practices, and Python automation tools the agent can invoke directly.
The senior ML engineer skill, for example, includes scripts for setting up MLOps infrastructure, running model deployment pipelines, and monitoring performance using MLflow. The senior architect skill ships an architecture diagram generator. The data engineer skill covers pipeline orchestration, data quality frameworks, and DataOps patterns. Beyond the engineering-team bundle, a separately maintained POWERFUL-tier engineering/ folder adds 78 advanced skills including agent designers, RAG architects, and a database designer with entity-relationship diagram generation.
The Adversarial Reviewer: Against the Self-Review Monoculture
One of the engineering bundle’s most distinctive entries is the adversarial-reviewer skill, designed to counter what its documentation calls “the self-review monoculture” — the tendency of AI coding assistants to approve their own suggestions when asked to review them.
The skill requires the agent to adopt multiple hostile reviewer personas in sequence, each with different priorities and failure modes. The documentation specifies that each persona must produce at least one finding; if a persona reports nothing wrong, it has not looked hard enough. Softening findings or hedging conclusions is explicitly prohibited. The skill’s recommended deployment contexts include security-sensitive code paths, pre-merge pull request reviews, and situations where the agent already indicated the code was fine.
Self-Improving Memory: How the Agent Curates Its Own Behavior
A second notable skill, self-improving-agent, extends Claude Code’s built-in session memory feature, which records project patterns to a MEMORY.md file. The base memory feature has no mechanism for evaluating which patterns deserve promotion to enforced rules, which have gone stale, or which recurring solutions should be extracted into reusable skills.
The self-improving-agent skill adds an analysis layer that can promote proven patterns from suggestions to hard rules and extract recurring solutions into new skill files — a mechanism for letting the agent curate its own long-term behavior through use rather than requiring manual curation by the developer.
How Does the Library Handle Installation?
For Claude Code users, the entire catalog is accessible through the native plugin marketplace with a single command. The engineering-team bundle is installable as a named domain package, as are marketing, product, compliance, C-level advisory, and eight other domains. For other agents, the repository’s conversion scripts handle the translation — a one-time setup after which the skills work natively in the agent’s expected directory structure.
Context efficiency gives the library a concrete economic case beyond convenience. Context engineering research has documented that a well-structured skills file that eliminates even five unnecessary clarification turns per session saves roughly 2,000 tokens. Over weeks of development on a token-priced API, that reduction compounds into measurable cost savings, and teams with multiple agents can apply the same skill library across all of them without reimplementing anything per tool.
Security Considerations Before Installing Any Skill Package
The broader AI agent skills ecosystem carries documented security risks that apply to any skill library. In January 2026, Cisco’s AI Defense team scanned the OpenClaw skills ecosystem and found that 26% of 31,000 skills analyzed contained at least one vulnerability. A coordinated campaign tracked as ClawHavoc placed 341 malicious skills on the ClawHub registry, including packages that used a fake “setup command” to download and execute credential-stealing malware.
Datadog Security Labs documented the Clawsights skill in the wild, which attempted to exfiltrate credentials from developers who installed it. Claude Code’s underlying model detected the malicious behavior and refused to execute it during Datadog’s testing — but the incident illustrates that installing any third-party skill package warrants review.
The claude-skills library ships a skill-security-auditor skill and runs a CI quality gate on every pull request. All 579 Python tools in the library use the standard library only, which limits the attack surface relative to packages that import external dependencies. For teams deploying any third-party skill library, reviewing the SKILL.md instructions and Python scripts before installation remains the first-line mitigation.
Beyond Engineering: the Rest of the Library
The 51-skill engineering bundle is one section of a library that now spans 17 domains and 345 packages. Other bundles include marketing (46 skills), a full C-suite advisory layer including CFO, CMO, CRO, COO, CHRO, CISO, and General Counsel personas each with dedicated forcing-question workflows, regulatory and quality management compliance, academic research, and enterprise research operations. All packages are available at github.com/alirezarezvani/claude-skills under the MIT license.
Frequently Asked Questions
What is the claude-skills library and how does it differ from Anthropic’s official Claude Skills?
The claude-skills library is a community-built, MIT-licensed collection of 345 skill packages maintained independently by developer Alireza Rezvani. It uses the same SKILL.md open specification that Anthropic published in December 2025, but it is not an Anthropic product and is not officially affiliated with Anthropic. The key distinction from most community libraries is its cross-agent design: the same packages work natively on 13 AI coding agents, with conversion scripts handling the per-agent format differences.
How does progressive disclosure loading work in AI coding agent skills?
When an agent starts a session, it reads only the name and a short description from each installed skill — typically a few dozen tokens per skill — and loads the full skill content only when a user request triggers it. This keeps context window usage low even when hundreds of skills are installed. The trigger is the description field in each SKILL.md, written specifically for the model with the phrases and contexts that indicate when the skill is relevant, not for human readers browsing a catalog.
Are AI agent skills safe to install from third-party libraries?
Skills run with the same permissions as the agent itself, which on most setups includes file system access and the ability to execute code. A malicious or compromised skill file can instruct the agent to run harmful commands. The claude-skills library ships a built-in security auditor skill and maintains a CI quality gate, and all its Python tools use the standard library only. For any third-party skill library, reviewing the SKILL.md instructions and Python scripts before installation is the recommended first step before deploying to a production environment.
What does it mean for the industry that AI coding agent skills are now cross-agent and standardized?
The SKILL.md specification is now supported by 32 tools including products from Google, Microsoft, JetBrains, and AWS. As the specification becomes a genuine cross-industry standard, the teams and libraries that define what domain-expert skill packages look like will shape how AI coding agents behave across the entire developer toolchain — regardless of which underlying model any individual tool uses. Open-source libraries competing to define what “senior engineer behavior” means in that landscape are building infrastructure that could outlast any individual agent product.