Blog · Zakhar Garan
How to save AI tokens
Working with AI agents every day, I noticed that what drives token usage most is not the model but how you frame the task. Here is what helps me spend less.
- The more detailed the prompt, the less time the AI spends hunting for details. If you know exactly where the fix goes, say so, so the agent does not waste tokens searching for files or functions.
- In Claude and Cursor the cache lives for 30–60 minutes, so it is better to finish a task in one go rather than come back to it hours later.
- New task, new chat. The previous context will not get in the way, and there is room for the new one.
- Simpler models for simple changes. You can ask the AI to delegate easy tasks to simpler subagents.
- Keep only short, genuinely necessary instructions in CLAUDE.md or AGENTS.md.
- Watch your spending: /usage and /context in Claude Code, /status in Codex.
- And it works surprisingly well to ask the AI itself how to save tokens: it will analyse your chats and give recommendations.