Claude code patterns

Seven battle-tested patterns from 2000+ hours of LLM development. Master the feedback loops that make AI reliable.

You own the failures. They come from bad prompts or bad context, so the fix is the system rather than the symptom.

Core Principle

"Any issue in LLM-generated code is solely due to YOU, the user." Problems trace back to two root causes: inadequate prompt engineering or insufficient context management.

// The patterns below fix both. No more blaming the model.

Core modules

Executable Skills

Performance Note

Models degrade ~50% at 50k tokens. Context is finite. Every token must earn its place. These patterns help you stay lean and effective across long sessions.

// Treat context like RAM. Clean it. Manage it. Respect its limits.

Looking for the strategy layer?

These patterns are the mechanics. Next level covers the decisions around them: how far to let an agent run unattended, how to stack independent reviews, and what counts as proof that a job is done.

Next level