Vous gaspillez 90% de vos tokens sans le savoir

Conversation Hygiene & Context Management
📌 Avoid long, continuous threads; every message sent includes the entire previous history, leading to context bloat and rapid token consumption. Use `/clear` to start fresh or initiate new conversations frequently.
📌 Monitor your context window usage. Once it reaches 40% capacity, performance may degrade; use `/compact` to summarize and retain only essential information.
📌 If using Claude Code (CLI), configure a status line to visualize real-time context usage, preventing you from "driving blind" and hitting limits unexpectedly.

Token Optimization & Model Selection
🤖 Choose the right model for the job: Use Claude 3.5 Opus for complex reasoning/structuring, Sonnet for execution and general tasks, and Haiku for simple operations like renaming files to save costs and tokens.
💬 Reduce verbosity to save on output tokens. Use tools or plugins like Caveman to enforce shorter, direct responses that consume fewer tokens than standard, conversational outputs.
⚠️ Avoid loading all MCPs (Model Context Protocol) by default. Check active context with `/context` and only enable specific MCPs or plugins when needed to prevent unnecessary resource drain.

Technical & Workflow Traps
📂 Avoid "obese" `claude.md` files; keep them concise, as this file is loaded at startup and consumes your context window before you even send your first message.
🚀 Prevent cache invalidation by avoiding model switches or adding MCPs mid-conversation; these actions force the model to re-process the entire chat history.
📑 Convert complex files (PDF, Word, Excel) to plain text or Markdown before uploading. Images and proprietary document formats can consume 15,000 to 20,000 tokens due to vision processing, compared to significantly fewer for text.
🧠 Use sub-agents only for isolated tasks (like specific searches). Running sub-agents on the main project context can cause the model to re-read files multiple times, exploding token usage.

Key Points & Insights
➡️ Understand the mechanism: Limits are triggered by token usage, not just message count. Every character—input and output—drains your allowance.
➡️ Strategic Scaling: While the $20/month plan is sufficient for light use, heavy power users should consider the $200/month tier for higher limits, as it is often more cost-effective than usage-based API billing for frequent workflows.
➡️ Efficiency Goal: The primary objective of these optimizations is not just cost-saving, but maintaining uninterrupted productivity throughout your workday. By avoiding even 3 of these 9 traps, you will significantly extend your daily usage limit.

📸 Video summarized with SummaryTube.com on Jun 02, 2026, 16:10 UTC