Skip to content

Skill Self-Creation and Self-Improvement

Skill Self-Creation and Self-Improvement refers to the capability of an AI system to autonomously generate new behavioral patterns ("skills") from successful task executions and refine those behaviors based on usage feedback^[001-TODO__Hermes_Agent_·_设计哲学与思维框架.md].

This concept moves beyond static, pre-programmed instructions or fixed prompts. It enables an agent to act as a Learning System that captures valuable experience into reusable code or structured prompts, ensuring that solutions to complex problems are retained and evolved over time^[001-TODO__Hermes_Agent_·_设计哲学与思维框架.md].

Core Concepts

This capability is often realized through a Closed Learning Loop, consisting of the following stages^[001-TODO__Hermes_Agent_·_设计哲学与思维框架.md]:

  1. Skill Self-Creation: The system analyzes complex tasks after completion. If the task was successful and appears to be reusable, the agent automatically synthesizes a new "skill" (a package of logic or prompt instructions) to capture that solution^[001-TODO__Hermes_Agent_·_设计哲学与思维框架.md].
  2. Skill Self-Improvement (Patching): When an existing skill is used and performs sub-optimally—or fails to handle a specific edge case—the system detects the shortcoming and applies a patch to update the skill's logic or prompt^[001-TODO__Hermes_Agent_·_设计哲学与思维框架.md].
  3. Knowledge Persistence (Nudging): To prevent valuable insights from being lost in transient chat logs, the system employs a "nudge" mechanism. This acts as an internal reminder or trigger to persist important discoveries or context into long-term memory (e.g., a vector store or file system) during or after a session^[001-TODO__Hermes_Agent_·_设计哲学与思维框架.md].

Technical Context

In implementations like the Hermes Agent, this capability is powered by a deep searchability of past conversations (e.g., using FTS5) and user modeling (e.g., Honcho Dialectic Modeling), which allows the system to contextualize new skills based on a deepening understanding of the user's specific needs and history^[001-TODO__Hermes_Agent_·_设计哲学与思维框架.md].

  • Closed Learning Loop
  • 20/80 Learning Principle: While 20/80 is a human learning strategy, skill self-creation is the AI equivalent—focusing on the high-impact patterns that solve the majority of problems.
  • [[失控感的自我強化迴路]]: In human behavior, lack of feedback leads to failure; in AI, closing the loop via self-improvement prevents the "loss of control" over code quality.

Sources

  • 001-TODO__Hermes_Agent_·_设计哲学与思维框架.md