How do I turn learned behaviors into permanent tools?

By using the /evolve command, the system identifies clusters of high-confidence instincts and generates the necessary markdown and configuration files to create new Skills, Commands, or Agents.

How does Continuous Learning v2 differ from standard Claude skills?

Unlike static skills, this system uses deterministic hooks to observe every interaction and builds a local library of behaviors that evolve and strengthen based on your actual usage patterns.

Does this impact the performance of Claude Code?

The observation hooks are lightweight shell scripts. The actual pattern analysis is performed by a background agent (Haiku), ensuring your main coding session remains fast and responsive.

Is my code data shared or uploaded anywhere?

No. All observations and instincts are stored locally in your homunculus directory. You have full control over what is processed, and only metadata 'instincts' are exported if you choose to share them.

What are 'Instincts' in this architecture?

Instincts are small, learned behaviors with metadata like confidence scores and evidence logs. They represent specific patterns, such as a preference for a code style or a specific testing workflow.

Continuous Learning v2

Name: Continuous Learning v2
Author: xu-xiang

byxu-xiang

•

생산성 및 워크플로

Transforms Claude Code sessions into a self-evolving library of automated behaviors and reusable skills through deterministic observation.

Continuous Learning v2 is an advanced architectural framework for Claude Code that captures patterns from your development sessions using high-reliability hooks. By observing user corrections, tool usage, and repeated workflows, the system generates 'Instincts'—atomic, confidence-weighted behaviors that evolve over time. These instincts are automatically clustered and transformed into specialized agents, custom commands, and reusable skills, allowing Claude to adapt to your specific coding style and project requirements without manual configuration.

주요 기능

01Deterministic session observation using PreToolUse and PostToolUse hooks for 100% capture rate.

0261 GitHub stars

03Background pattern detection powered by Claude Haiku to minimize workflow interruption.

04Local-first storage and privacy-focused export/import system for sharing learned patterns.

05Atomic Instinct Model with weighted confidence scoring (0.3 to 0.9) to represent learned behaviors.

06Automated evolution of high-confidence instincts into permanent agents, commands, or skills.

사용 사례

01Automatically codifying team-specific coding standards and architectural patterns through daily interaction.

02Capturing complex, multi-step debugging workflows to create one-click resolution commands.

03Building project-specific 'specialist' agents that learn from developer corrections and repeated tool sequences.

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add xu-xiang/everything-claude-code-zh continuous-learning-v2

For use in Claude.ai and ChatGPT

Download Skill