Core Principles

Progressive disclosure. Load only what is needed, when it is needed. Keep global context small. Push specificity downward into rules, skills, agents, and teams. Root context is a behavioral contract, not a tutorial. Every line must change behavior.

See references/principles.md for the full rationale.

Agent System Model

Five layers, smallest scope first:

1. Path-scoped rules (.claude/rules/) — passive guidance for specific files
2. Skills (.claude/skills/) — repeatable workflows loaded on demand
3. Subagents (.claude/agents/) — specialist workers with isolated context
4. Agent teams — multiple independent sessions coordinated by a lead
5. Hooks/settings — deterministic enforcement at lifecycle boundaries