continuous-learning-v2

Captures Claude Code session activity via hooks and evolves observed patterns into reusable instincts, skills, and commands.

2|Updated Feb 25, 2026
One-click install
npx skills add https://github.com/adamreger/ecc-antigravity --skill continuous-learning-v2-adamreger
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: continuous-learning-v2
Source: https://github.com/adamreger/ecc-antigravity/tree/main/skills/continuous-learning-v2
Command: npx skills add https://github.com/adamreger/ecc-antigravity --skill continuous-learning-v2-adamreger

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) components.

What problem does it solve? Valuable patterns from your Claude Code sessions—corrections, error fixes, repeated workflows—are lost when sessions end. This Skill observes tool usage deterministically via hooks, distills patterns into atomic confidence-scored instincts, and evolves them into reusable skills, commands, and agents. ## Core Features & Use Cases - Hook-Based Observation: PreToolUse/PostToolUse hooks log every tool call to a local JSONL observations file, capturing 100% of session activity instead of relying on probabilistic skill activation. - Instinct Management CLI: A Python CLI provides status, import, export, and evolve commands to review confidence-scored instincts, share them with teammates, and cluster them into skills, commands, or agents. - Background Observer Agent: A Haiku-based background agent periodically analyzes observations to detect user corrections, error resolutions, repeated workflows, and tool preferences. - Use Case: After several sessions where you repeatedly correct Claude to use functional patterns over classes, the system creates a 'prefer-functional-style' instinct with rising confidence, which you later evolve into a team-shared coding-style skill. ## Quick Start Set up continuous learning by adding the observation hooks to my Claude settings and show me the status of any learned instincts.

Frequently Asked Questions about continuous-learning-v2

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I set up automatic learning from Claude Code sessions?▼

Add the observe.sh script as PreToolUse and PostToolUse hooks in your ~/.claude/settings.json, pointing to the skill's hooks directory. The hooks then log every tool call to ~/.claude/homunculus/observations.jsonl for pattern analysis.

How do I export and import learned instincts between machines?▼

Use the instinct CLI: run 'instinct-cli.py export -o file.yaml' to export instincts with optional domain and confidence filters, and 'instinct-cli.py import <file-or-url>' on another machine to import them into the inherited instincts directory.

What is the difference between continuous-learning v1 and v2?▼

v1 used a Stop hook and full-skill generation, while v2 uses PreToolUse/PostToolUse hooks for deterministic capture, atomic confidence-scored instincts, and an evolution pipeline that clusters instincts into skills, commands, or agents. v2 remains backward compatible with v1 learned skills.

Does the observation hook send my code or conversations to external servers?▼

No. Observations stay local in ~/.claude/homunculus/observations.jsonl, and only abstracted instinct patterns can be exported. The documentation states no actual code or conversation content is shared.

Why is the observer agent not creating any instincts?▼

The observer is disabled by default in config.json and only analyzes when at least 10-20 observations exist. Enable it in config, start it with start-observer.sh, and ensure the claude CLI is available for the Haiku analysis step.