Why Your Claude Setup Is Burning Tokens
Most Claude Code users create an AGENTS.md or CLAUDE.md file. A thousand lines is not unusual. They fill it with everything they want Claude to remember: - tech stack - coding conventions - tone of voice - folder structure - review checklist - deployment process - project gotchas The problem? That file loads into the context window on every single turn. If it’s 7,000 tokens, you pay 7,000 tokens just to ask Claude what file to edit. You pay it again when you ask a follow-up. You pay it again when Claude replies. By turn twenty, you may have burned 100,000+ tokens on instructions that were only relevant for maybe three of those turns. Two things break: 1. CostTokens are not free. You’re paying for context you are not using. 2. QualityYou hit the context-fill danger zone sooner. The model has less room for the actual task, and performance starts to degrade. My hot take: Stop putting everything in one giant always-loaded instruction file. Use skills instead. I tested one skill file: - Full skill body: 944 tokens - Name + description only: 53 tokens That’s an 18x difference. And that difference compounds on every turn, every session, for every user on your team. The better pattern is simple: Keep the always-loaded context small. Load the detailed instructions only when the task actually needs them. Curious: how big is your AGENTS.md or CLAUDE.md file right now?