Show HN: A governance pattern for self-evolving AI skills

In a significant advancement for AI skills development, a new design pattern has emerged for Claude Code Skills that allows these systems to improve their accuracy and efficiency over time. This self-evolving capability is particularly beneficial in domains where AI can continuously learn and adapt based on user interactions, such as database investigation and codebase analysis. Unlike traditional static skills, which are limited to a one-time setup, this innovative pattern enables AI to accumulate valuable domain knowledge during use, thereby enhancing its effectiveness without unnecessary complexity or bloating.

Understanding Self-Evolving Skills

The concept of self-evolving skills is grounded in the framework of Inter-test-time Context Evolution with Text-Feedback Governance, as outlined in the academic literature by Gao et al. (2026) in their survey of self-evolving agents. Traditional skills are static; they are created by an author, invoked by users, and do not evolve or grow in knowledge. In contrast, the self-evolving skills pattern allows an AI to learn from each interaction, capturing insights such as table relationships, query patterns, and business rules that can be beneficial for future tasks.

The reference implementation of this pattern is a Self-Evolving Skill specifically designed for MySQL database investigation. Users can install this skill to observe its functionality firsthand. The installation process includes an interactive wizard that collects MySQL connection details, tests the connection, and initializes the knowledge system. Users can also initiate a conversation with the AI to ask database-related questions, prompting the skill to activate and begin evolving its domain knowledge through the interaction.

This self-evolving mechanism is crucial because it prevents the knowledge base from deteriorating into irrelevant information. The Five Gates protocol, a key component of this design, primarily functions to filter out unworthy knowledge, ensuring that only valuable insights are retained. Most interactions will not yield knowledge worth storing, and the protocol's main role is to reject unnecessary additions rather than accept them.

The Mechanics of Knowledge Evolution

A significant distinction between traditional skills and self-evolving skills lies in their respective levels of knowledge management. In a traditional skill, Level 3 is static reference documentation, while in a self-evolving skill, Level 3 represents a dynamic knowledge base that is actively maintained by the AI under a governance protocol. This allows for a more nuanced and responsive approach to knowledge management.

The complexity of the confidence decay model, which incorporates elements such as exponential decay and Bayesian feedback adjustment, is designed to be manageable for the AI. The pattern effectively separates concerns, allowing for more intricate formulas without overwhelming the skill's documentation. This separation also ensures that each formula can be unit-tested, enhancing the reliability of the system.

An interactive decay model visualization allows users to explore how various parameters affect the confidence formula. This tool provides insights into how knowledge types, feedback, and time influence decisions regarding trust, verification, and revalidation.

The design pattern also addresses the limitations of context windows in AI systems. Since a skill acts as a prompt injection layer, it is essential to manage the context window effectively. The routing table ensures that the injection size remains controlled, regardless of the knowledge base's growth, preventing unnecessary information from overwhelming the system.

Implications for Creators and Technologists

The introduction of the Self-Evolving Skill design pattern has profound implications for creators and technologists. It represents a shift towards more adaptive and intelligent AI systems capable of evolving in response to real-world interactions rather than relying on pre-set knowledge. This evolution is particularly relevant in fields that require continuous learning and adaptation, such as data management and business intelligence.

The pattern emphasizes the importance of user-driven knowledge accumulation. Changes to the knowledge base are only triggered by genuine user interactions, ensuring that the AI's growth aligns with actual business needs. This contrasts with previous academic systems that explored autonomy in learning, which may not always align with practical applications.

Moreover, the stability-plasticity dilemma is addressed within this framework. A mature skill that ceases to grow after covering the majority of everyday scenarios is considered healthy, indicating that the system has reached a point of stability. This stability is crucial to prevent catastrophic forgetting, a common issue in systems that undergo constant change.

The safety advantages of operating at the context layer rather than modifying model parameters further enhance the reliability of self-evolving skills. Even in cases where the knowledge base develops errors, the impact is limited to incorrect references rather than fundamental changes in behavior.

In summary, the Self-Evolving Skill design pattern not only advances the capabilities of AI systems but also provides a structured approach to managing knowledge evolution in a way that is both efficient and safe. As this pattern evolves from a database-specific application to a more general-purpose governance mechanism, it opens new avenues for innovation in AI development across various domains.