A free, MIT-licensed library called claude-skills has grown to 345 production-ready skill packages that work without modification across 13 AI coding agents, including Claude Code, OpenAI Codex, Gemini CLI, Cursor, Hermes Agent, and Windsurf. Developer Alireza Rezvani updated the library to v2.10.3 on June 10, according to TechTimes.
The library addresses a problem that compounds as engineering teams adopt multiple AI coding tools simultaneously: there is no standard way to make different agents behave like domain experts using the same set of instructions. A 2026 DX Report survey found that 73% of engineering leaders cited inconsistent behavior across their team’s AI tools as a primary productivity problem, according to TechTimes.
How the Cross-Agent Format Works
Every package in claude-skills follows Anthropic’s SKILL.md open specification, published in December 2025. A skill is a folder containing instruction files, reference guides, and Python automation scripts. The loading mechanism uses progressive disclosure: agents scan each skill’s name and short description at session start (a few dozen tokens each) and load full skill content only when a user’s request triggers it. This means teams can install hundreds of skills without flooding the model’s context window.
Claude Code, Codex, and Gemini CLI each read skill files from different directory paths. The repository ships conversion scripts that translate the Claude Code-native layout into each agent’s expected structure, so a team can define one authoritative set of engineering behaviors and propagate them across every tool in use.
All 579 Python tools bundled with the library run on the standard library alone: no pip installs, no external API keys, no LLM calls. That constraint eliminates supply chain dependencies and ensures the scripts run anywhere Python runs.
51 Engineering Personas and an Adversarial Reviewer
The core engineering-team bundle covers 51 roles: senior architect, frontend, backend, fullstack, QA, DevOps, SecOps, data engineer, ML engineer, prompt engineer, and computer vision specialist, among others. A separately maintained advanced tier adds 78 more skills including agent designers, RAG architects, and database designers with entity-relationship diagram generation.
One notable entry is the adversarial-reviewer skill, designed to counter what its documentation calls “the self-review monoculture,” the tendency of AI coding agents to approve their own output. The skill forces the agent to adopt multiple hostile reviewer personas in sequence, each required to produce at least one finding. Softening conclusions is explicitly prohibited.
A second skill, self-improving-agent, extends Claude Code’s session memory by adding an analysis layer that can promote proven patterns from suggestions to enforced rules and extract recurring solutions into new skill files automatically.
Security Context
The broader agent skill ecosystem carries documented risks. In January 2026, Cisco’s AI Defense team scanned the OpenClaw skills marketplace and found that 26% of 31,000 skills analyzed contained at least one vulnerability, according to TechTimes. The claude-skills library ships a skill-security-auditor package and runs a CI quality gate on every pull request as mitigations.
Beyond engineering, the library spans 17 domains and includes marketing (46 skills), C-suite advisory personas, regulatory compliance, and enterprise research operations. The full catalog is available on GitHub under the MIT license.