How to Save Claude Code Tokens: Prompts That Keep Claude Code Short and Focused
Claude Code can waste tokens surprisingly easily.
You ask it to change one small thing, and instead of simply making the change, it explains the project, describes what it found, discusses several possible approaches, repeats your requirements, and gives you a long summary.
That extra output may sound helpful, but when you are using Claude Code heavily, it can become expensive and slow.
The solution is simple: tell Claude Code exactly how much information you want and what it should NOT explain.
You don't need a complicated prompt.
You need a short instruction that controls the behavior.
The Core Prompt
If you want Claude Code to focus on coding instead of writing explanations, use this:
Work directly on the requested task.
Be concise.
Do not explain information I did not ask for.
Do not repeat my instructions.
Do not describe obvious changes.
Do not provide long summaries.
Only mention important changes, errors, blockers, or decisions.
If the task is clear, execute it without asking unnecessary questions.
This tells Claude Code something important:
Your goal is the code, not the conversation.
The Ultra-Short Version
If you want to save even more tokens, use:
Execute the task directly.
Be extremely concise.
No unnecessary explanations.
No repetition.
No long summary.
Only report:
- what changed
- important errors
- blockers
- required next step
This is useful for repetitive coding work where you already understand your project.
Tell Claude Code What NOT to Return
One of the easiest ways to reduce output is to explicitly remove unnecessary response types.
For example:
Do not:
- repeat my request
- explain basic concepts
- describe every file you inspected
- give unnecessary alternatives
- write a long summary
- explain code that is self-explanatory
- include information unrelated to the requested task
Only provide information necessary to complete or verify the task.
This can be especially useful when Claude Code is working inside a large project.
Instead of asking for a report about everything it did, ask for only the result.
Use a Token-Saving Coding Prompt
Here is a more complete reusable prompt:
You are working as my coding agent.
Optimize your responses for low token usage.
Rules:
1. Execute clear tasks directly.
2. Do not repeat my instructions.
3. Do not explain obvious code.
4. Do not provide unnecessary background information.
5. Do not list files unless relevant.
6. Do not provide multiple solutions unless the first solution cannot work.
7. Prefer the smallest correct change.
8. Keep explanations to 1–3 short sentences when necessary.
9. Report only important errors, blockers, decisions, and results.
10. If the task succeeds, give a very short completion message.
Prioritize:
correctness > simplicity > explanation length.
Give Claude Code the Minimum Context
Another major source of wasted tokens is giving an AI agent information that has nothing to do with the current task.
For example, if you want to change a button color, you don't need to explain your entire business, marketing strategy, company history, or every feature in the application.
Instead:
Change the primary button color to #111111.
Do not modify layout, spacing, typography, functionality, or other components.
Make only the required change.
The instruction is narrow.
The output can therefore be narrow too.
Use “Only Change What Is Necessary”
This is one of the most useful instructions for coding agents.
Make the smallest possible change that solves the problem.
