A two-layer extensibility pattern that makes Claude Code progressively self-correcting: hooks for real-time session feedback, skills for periodic infrastructure auditing (Luca Dellanna, Feb 2026). ## Layer 1 — Real-Time Hooks (intra-session) - **Self-improvement injection**: a `userPromptSubmit` hook that detects sessions with 8+ tool calls and appends one optimization hint (reusable skill, memory pattern, workflow fix); skips exploratory sessions. - **Conditional prompt linker**: appends a clarity check on long prompts, asking Claude to confirm the desired outcome. ## Layer 2 — Periodic Audit Skills (cross-session) Skills audit (line counts + review prompts), CLAUDE.md audit (redundancy/verbosity), and reflect (extract learnings from a conversation) automate the "assess" and "compound" steps of the improvement loop. ## Design Principle Trigger accuracy over hint content: a hook that fires on the wrong task type burns more tokens than no hook. Use script-based filtering, conditional skipping, and tool-call/word-count thresholds as activation gates. ## Related Concepts - [[Compound Engineering Method]] — the macro loop this mechanizes - [[Self-Improving Scheduled Agent Pattern]] - [[Model-Invoked vs User-Invoked Skills Architecture]] --- *Source: Luca Dellanna — [x.com/DellAnnaLuca](https://x.com/DellAnnaLuca)*