Self-Improving Agent

An always-on observer that captures user corrections and preferences, tracks repetition across projects, and escalates behavioral rules to the appropriate AGENTS.md file.

Two-Tier Activation

Tier 1: Observation (always in context)

The YAML description above is loaded every session. It contains everything needed for the lightweight per-message check:

1. After every user message, silently evaluate: did the user correct me, express a preference, or show frustration about repetition?
2. If yes, run one CLI command to record it.
3. Do NOT load the rest of this SKILL.md for observation alone.

Tier 2: Escalation & Writing (loaded on demand)