Cognitive Core Skills: A Universal Taxonomy for AI Capability Assessment
Cognitive Core Skills is a structured, industry-neutral taxonomy designed to evaluate and benchmark the cognitive capabilities of large language models (LLMs), small language models (SLMs), AI agents, and world models. The taxonomy categorizes essential mental operating capabilities into eight core domains: perception, memory, reasoning, planning, action, verification, learning, and governance.
The taxonomy is not limited to text-based interactions. It also includes human-action cognitive skills and governance-oriented constructs, ensuring a comprehensive framework for assessing AI systems in real-world workflows. This distinction is critical because while LLMs excel in language understanding and synthesis, they lack inherent continuity, autonomous action, or built-in safety mechanisms.
The taxonomy is implemented as a modular package with machine-readable schemas, validation tests, and continuous integration (CI) pipelines. It is structured to support interoperability across different AI system types, including those with constrained inference (SLMs) or simulation-based forecasting (world models). Each capability is documented with clear definitions, benchmarks, and evaluation rubrics, making it adaptable for research, development, and operational deployment.
The initial release (version 1.0.0) includes 159 skill entries across 13 domains, covering both text-based and human-action-oriented competencies. These skills are organized into JSON and YAML formats for easy integration into existing workflows, while markdown skill cards provide human-readable documentation. The taxonomy also introduces governance-focused constructs, such as verification protocols and cognitive debt indexing, to address safety and reliability concerns in AI-driven systems.
The project is licensed under the MIT License, ensuring permissive reuse and modification.
How It Works
The Cognitive Core Skills taxonomy operates as a modular framework composed of three primary components: schemas, skill cards, and benchmarks. The schemas define the structural requirements for each cognitive skill, ensuring consistency across representations in JSON, YAML, and markdown formats. Skill cards provide detailed descriptions of individual capabilities, including their purpose, related skills, and evaluation criteria. These cards are automatically generated from the taxonomy data to maintain synchronization.
Benchmarks are integrated into the taxonomy to assess AI system performance against predefined rubrics. These rubrics evaluate skills such as long-horizon planning, consequence understanding for human actions, and verification of execution outcomes. The benchmarks are stored in machine-readable formats (e.g., rubric-task-traces.json) and are validated through automated tests to ensure reliability. Continuous integration pipelines further guarantee that updates to the taxonomy or skill definitions do not introduce inconsistencies.
For example, LLMs are strong in language understanding and synthesis but lack continuity and autonomous action. SLMs, optimized for low-latency and private inference, may struggle with open-domain depth. AI agents, which combine LLMs with memory, planning, and tool integration, require robust verification and control mechanisms to operate safely.
To maintain alignment with practical use cases, the taxonomy includes specialized domains such as LLM Wiki operations (persistent knowledge management) and OKF-style knowledge cataloging (structured knowledge organization).
Use Cases
The Cognitive Core Skills taxonomy is primarily intended for testing and red teaming AI systems, but its applications extend to capability design, benchmarking, and operational safety assessments. Developers and researchers can use the taxonomy to identify gaps in AI agent functionality, such as weaknesses in long-horizon planning or verification protocols. By mapping system capabilities against the taxonomy, teams can prioritize improvements in critical areas such as governance, memory continuity, or action selection.
In red teaming exercises, the taxonomy provides a structured approach to evaluating AI systems for unintended behaviors or safety vulnerabilities. This capability is particularly valuable in high-stakes environments, such as autonomous systems or AI-driven decision-making tools.
The taxonomy also supports the development of AI agents for specific workflows by offering modular, reusable skill definitions. Organizations can adapt the taxonomy to their unique requirements, integrating it with existing agent frameworks or toolchains. The machine-readable formats (JSON, YAML) and automated validation ensure seamless incorporation into CI/CD pipelines, while the governance-oriented constructs provide a foundation for compliance and risk management.
For world models and simulation-based AI systems, the taxonomy offers a framework for evaluating state modeling, forecasting accuracy, and dynamic adaptation.
Key Features and Capabilities
The taxonomy includes 159 skill cards, each representing a distinct cognitive capability required for AI systems to function autonomously in real-world scenarios. These skills are organized into 13 domains, covering text-based interactions, human-action competencies, and governance mechanisms. Each skill card contains metadata such as unique identifiers, related skills, and evaluation criteria, enabling systematic assessment and comparison across AI models.
One of the taxonomy’s standout features is its support for governance-oriented constructs, including verification protocols and cognitive debt indexing.
The taxonomy is designed for interoperability, with schemas that define the structure of skill definitions in JSON and YAML formats. These schemas ensure consistency across implementations, while the markdown skill cards offer human-readable documentation for developers and researchers. The inclusion of validation tests and CI pipelines guarantees that updates to the taxonomy remain reliable and error-free.
For developers looking to extend the taxonomy, the project provides clear guidelines for adding new skills. The process involves updating the JSON and YAML data files, regenerating markdown skill cards, and running validation tests to maintain synchronization.
These benchmarks cover a range of cognitive skills, from basic reasoning to complex long-horizon planning.
Recent updates to the taxonomy have introduced specialized skills such as human-action consequence understanding and long-horizon planning. These additions reflect the growing need for AI systems to operate safely and effectively in dynamic, real-world workflows.