How does the system handle data privacy?

All observations and learned instincts are stored locally on your machine within the ~/.codex/mdt/ directory, ensuring your coding patterns and project data remain private.

What is an 'instinct' in this system?

An instinct is a small, learned behavior derived from repeated patterns in your workflow, weighted by a confidence score based on the frequency and consistency of observed evidence.

What happens when an instinct reaches high confidence?

High-confidence instincts can be 'promoted' or evolved into permanent, tool-facing skills, commands, or agents that Claude can use natively.

Can I use these learned patterns across different projects?

Yes. The system supports project-scoped instincts for repository-specific rules and global instincts for general best practices that apply to all your work.

Is the learning process automatic or manual?

It follows a manual-first baseline where you explicitly capture and analyze summaries, though an optional background observer can be enabled for automated analysis.

AI Learning & Instincts

Name: AI Learning & Instincts
Author: flatrick

byflatrick

0•

生产力与工作流

Captures and evolves repetitive developer patterns into reusable AI instincts and personalized skills.

AI Learning & Instincts is a specialized performance optimization system for the MDT framework that allows Claude to learn from your specific coding habits and project-level preferences. By tracking observations and assigning confidence scores to patterns—such as preferred code styles or common multi-step workflows—it transforms repetitive behaviors into "instincts." This ensures your AI assistant becomes increasingly personalized over time, moving from tentative suggestions to core behaviors while maintaining strict project isolation or promoting global best practices when patterns recur across multiple repositories.

主要功能

010 GitHub stars

02Automated weekly retrospectives to identify automation and MCP integration candidates

03Instinct-based learning with confidence-weighted evidence scoring (0.3 to 0.9)

04Evolution paths to promote high-confidence instincts into permanent skills or commands

05Project-scoped storage to prevent pattern contamination between different codebases

06Manual-first capture baseline with an optional opt-in background observer

使用场景

01Identifying repeated shell commands to convert them into automated scripts or custom tools

02Personalizing AI behavior across diverse repositories while maintaining local security boundaries

03Enforcing project-specific coding conventions like React hook patterns or error handling strategies

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add flatrick/mdt ai-learning

For use in Claude.ai and ChatGPT