Optimizing token consumption is crucial for maintaining fast and accurate AI coding assistants. Managing tokens effectively prevents context bloat, which leads to increased latency, instruction forgetting, and hallucinations. Implementing structured habits ensures a productive feedback loop. Start with balanced models and scale up only when necessary, utilizing reusable skills from the outset to keep prompts concise. Automate repetitive tasks with scripts and CLI tools, and delegate computationally intensive jobs to sub-agents. Divide complex projects using high-reasoning sessions for planning and low-token sessions for execution. Shift verification left by automating testing early and reserving expensive checks for later stages. Use the undo function to correct a drifting agent rather than piling on corrective prompts. Be specific with context, pointing the agent to precise locations for fixes. Iterate on rules by updating AGENTS.md or skill files instead of repeating corrections. Avoid uncontrolled loops by setting strict limits and using event-driven wakeups instead of polling. Finally, start new sessions for distinct topics to ensure the AI has only the necessary context.