Self-Improving Skills in Claude Code

Key Takeaways

• CLAUDE.md as living memory — Claude Code is stateless between sessions. The CLAUDE.md file persists knowledge across conversations. When Claude makes a mistake, adding the pattern to CLAUDE.md prevents repetition.
• One sentence triggers the cycle — When Claude errs, instead of just fixing it, prompt Claude to reflect on the mistake, abstract the general pattern, and document it. This turns individual fixes into systematic improvements.
• Compounding knowledge — Boris Cherny's team at Anthropic adds mistakes to CLAUDE.md multiple times per week. Every team member contributes. The codebase becomes a "self-correcting organism."
• Self-written code — Anthropic reports that 80-90% of Claude Code is now written by Claude Code itself. The self-improvement loop extends to the tool building itself.

The Self-Improvement Pattern

1. Claude makes a mistake
2. Instead of just fixing, prompt Claude to analyze the error
3. Claude extracts the general pattern
4. The pattern gets added to CLAUDE.md
5. Next session, Claude avoids the mistake automatically

Connections

• boris-cherny-on-what-grew-his-career-and-building-at-anthropic — Boris Cherny describes this exact CLAUDE.md philosophy from the Claude Code team's daily practice
• writing-a-good-claude-md — Practical guidelines for crafting effective CLAUDE.md files that maximize Claude's instruction-following
• claude-code-skills — Skills extend Claude's capabilities similarly, but activate automatically based on semantic matching rather than explicit instruction