Get Ahead of 99% of Claude Code Users

After 800+ hours in Claude Code, Artem Zhutov shares what separates power users from everyone else: feedback loops for self-improvement, deep observability, and treating personal notes as executable context.

Key Takeaways

• Analyze your conversations — All Claude Code sessions are stored as JSON files locally. Build a skill that parses them to find friction points (repeated corrections, ignored instructions) and automatically suggests CLAUDE.md improvements.
• Verbose output reveals the black box — Enable verbose output in settings to see exactly what Claude reads, what prompts run, and how the system prompt is structured. Understanding the internals helps you work within the constraints.
• Most skills go unused — Having 35 skills means nothing if you only use 5. Curate ruthlessly. Use Obsidian as a dashboard to track skill status (active/archived) and surface forgotten ones.
• MCP tools waste context — Loading MCP definitions (Notion, Linear) consumes 14% of your context window before you even start. CLI tools loaded on-demand via skills are more efficient.
• Notes become executable — Store goals, tasks, and context in a structured vault. Claude Code transforms static notes into an active system that can read your goals, check your tasks, and execute against your personal context.
• Session handoffs matter — When context runs low, don't rely on default compaction. Document current progress in a file Claude can reload in the next session, tagged for easy retrieval.

Workflow: Session Analysis Feedback Loop

flowchart LR
    Sessions[JSON Conversation Files] --> Analyze[Session Analysis Skill]
    Analyze --> Friction[Find Friction Patterns]
    Friction --> Suggest[Suggest CLAUDE.md Rules]
    Suggest --> Memory[Update Agent Memory]
    Memory --> Sessions

Steps Summary

1. Enable verbose output — Settings → search "verbose" → turn on
2. Build session analysis skill — Parse local JSON files for friction indicators (keywords like "wrong", "not what I asked")
3. Set up Obsidian dashboard — Track skills with status, surface usage patterns via Dataview
4. Replace MCP with CLI skills — Load external integrations on-demand rather than upfront
5. Store personal context — Goals, tasks, routines in structured markdown for Claude to act on
6. Create handoff workflow — When context fills, document progress and tag for next session

Notable Quotes

"All of your data, all your conversations is here on your computer and you can use it for improving how you work."

"English right now is a programming language... the notes aren't static, they're useful context which the agent can act upon."

"MCP tools fill up 14% of our conversation which we could have spent on interacting with the agent."

Connections

• self-improving-skills-in-claude-code — Same feedback loop pattern: capture mistakes in CLAUDE.md to prevent repetition across sessions
• obsidian-claude-code-workflows — Community patterns for combining Obsidian with Claude Code, including master context files and batch operations
• claude-code-skills — Official documentation on skills, the mechanism Artem uses to load CLI tools on-demand instead of MCP