11 Token-Saving Tips for AI-Assisted Coding

Tokens aren’t infinite—behind every LLM call is a real machine doing work. If you’re using AI coding assistants, you’ve probably felt the pain of context bloat: slower responses, forgotten instructions, and a fatter bill. But with a few smart habits, you can keep your feedback loop fast, precise, and productive. Here are 11 principles to help you spend tokens wisely.

1. Start with a balanced model

When in doubt, begin with the default Gemini 3.5 Flash (Medium reasoning). Gauge complexity as you go. Scale up to larger models or higher reasoning only if a task fails or needs complex design.

2. Use skills from the beginning

Don’t explain your workflow or testing rules in every prompt. Package reusable skills with SKILL.md files and scripts. The agent triggers them automatically, keeping prompts clean and avoiding unnecessary searches.

3. Automate with scripts and CLI tools

For repetitive chores like formatting files or extracting logs, have the agent create simple local tools. Use official CLI tools for setup, linting, and testing. Run read-only commands to research the codebase before writing code—avoid long trial-and-error loops.

4. Delegate output-heavy tasks

Delegate deep research or separating frontend/backend work to sub-agents. Once they’re done, you only reconcile the final results, not the full trajectory.

5. Divide and conquer

Use a high-reasoning, long-context session (“Elephant”) to generate a detailed execution plan (“Goldfish”). Execute that plan in a clean, low-token session. Checkpoint progress often with commits or artifacts so you can restart from a clean state when context fills up.

6. Shift verification left

Automate testing early. Run local builds and unit/functional tests before UI testing. Tell the agent to perform expensive smoke-tests in the browser right before handoff. Save expensive verification loops for the end of the milestone.

7. Undo when adrift

If the agent drifts and you know the fix, use the Undo button or revert your files. Don’t pile corrective prompts on a broken state—that poisons the context.

8. Be specific with context

A clear instruction with a few spelling errors is better than a grammatically accurate broad request. Point the agent to the exact file, section, or error with an obvious annotation like // SHOULD BE X, NOT Y, FIX THIS. Use inline comments so the agent knows exactly where you want the fix.

9. Iterate on rules

If you keep correcting the agent’s behavior, update your global rules in AGENTS.md or edit the skill. Fix the instructions instead of prompting repeatedly—so the change persists.

10. Avoid uncontrolled loops

Supervisor loops that scan projects for pending work can find optimizations, but they can burn your entire token budget. Set strict limits and stop conditions. Use event-driven wakeups instead of polling.

11. Start new sessions for each new topic

If you’re continuing on the same topic, reuse the chat to keep context. But if you change topics, start a new chat. The agent will provide better answers with fewer tokens if it only pulls in the context it needs.

Token optimization is about directing the AI’s attention. By using a tiered approach, you keep development fast and output sharp while optimizing spending. Happy coding!

Original announcement published on Google Cloud.