CLAUDE.md and SKILL.md · Master AI Automation

CLAUDE.md and SKILL.md

"Why did Claude forget my project rules again?"

"Why is my token cost skyrocketing on simple tasks?"

If you've asked yourself this while using Claude Code, you likely need to optimize your memory architecture.

Anthropic splits agent memory into two distinct layers to solve this exact problem:

1️⃣ The "Always On" Layer (CLAUDE.md): Automatically loaded every session. It acts as the guardrails for your repo, holding your TypeScript strict preferences, DB versions, and testing requirements.

2️⃣ The "On Demand" Layer (SKILL.md): Only invoked when a specific task requires it. Perfect for keeping heavy, repetitive instructions (like production deployment checklists) out of your daily conversation flow until they are actually needed.

Comment 'Guide' and I will send the handy cheat sheet summarizing the differences, token limits, and implementation examples.

How are you structuring your AI agent prompts? Drop your tips below!

5 comments