continuous-learning-v2

You can automatically capture Claude Code session observations using hooks, which deterministically record workflows and tool usage, then convert them into atomic instincts with confidence scoring that evolve into skills, commands, or agents.

Atomic instincts with confidence scoring are granular behavioral patterns extracted from session observations, quantified by reliability to determine if they should remain project-scoped or be promoted globally to evolved capabilities like skills and agents.

To isolate learning instincts per project and prevent cross-project contamination, apply project-scoped instincts by default, keeping patterns separated per project while allowing universal patterns to be promoted to a global layer.

No external dependencies are required to process session hooks, as the system relies entirely on internal scripts to capture tool usage and manage the evolution pathway from instincts to skills, commands, and agents.

The best way to evolve observed workflows into agents and commands is through a deterministic observation pipeline that clusters atomic instincts by confidence scoring, promoting universal patterns from a project scope to a global layer.