How to Reduce Claude Code Token Usage with Caveman: A Complete Setup Guide

If you use Claude Code regularly, you've probably noticed one problem: Claude can be too verbose.

For a simple task like:

Fix the authentication bug.

Claude may explain the issue, describe its plan, provide progress updates, and finish with a long summary - even when you only need the result.
That's where Caveman comes in.

Caveman is an open-source skill/plugin that makes AI coding agents communicate more concisely while preserving important technical information such as code, commands, file paths, and errors.

Instead of:

I've identified the issue with the authentication middleware. I'll update the implementation and run the relevant tests."

You get something closer to:

Auth middleware issue. Fix implementation. Run tests.

Same work. Fewer unnecessary words.

Why Use Caveman?

A Claude Code session can involve prompts, repository context, files, tool results, conversation history, and model responses. As the session grows, so does the context.

Caveman primarily helps by reducing output verbosity. Newer Caveman Proxy functionality can also compress some input/context before it reaches the model.

Important: A claimed 65% reduction mainly refers to output tokens. It does not automatically mean your total Claude bill will be 65% lower.

Your actual savings depend on your workload, context size, tool calls, caching, reasoning, and number of iterations.

Install Caveman with Claude Code

Prerequisites

You need:

  • Claude Code
  • Node.js 18+
  • npm
  • macOS, Linux, WSL, or supported Windows environment

Check your installation:

node --version
npm --version
claude --version

Install the Caveman plugin

Add the marketplace:

claude plugin marketplace add JuliusBrussee/caveman

Then install Caveman:

claude plugin install caveman@caveman

Start Claude Code:

claude

Then activate Caveman:

/caveman

You can select a compression level:

/caveman lite
/caveman full
/caveman ultra

The available modes can vary depending on the installed version.

Which Mode Should You Use?

My recommended setup:

TaskMode
Normal developmentfull
Debuggingfull
Running testsfull
Repetitive tasksultra
LearningOff
DocumentationOff
Architecture discussionsOff

Aggressive compression isn't ideal when you need detailed explanations, documentation, tutorials, or architectural discussions.

Measure Your Actual Savings

Don't rely only on the headline percentage.

Run the same type of workload before and after enabling Caveman and compare:

  • Input tokens
  • Output tokens
  • Total tokens
  • Number of turns
  • Completion time

You can also check Caveman's statistics:

/caveman-stats

This lets you see estimated and session-level savings rather than simply assuming Caveman is helping.

Compress CLAUDE.md

Caveman can also compress context files such as:

CLAUDE.md

using:

/caveman-compress <file>

This is useful when your CLAUDE.md contains a lot of repetitive instructions. The project says its compression preserves important elements such as code, URLs, and paths.

Caveman vs Normal Claude Code

The basic idea is simple:

Normal Claude Code
User → Claude Code → Claude → Detailed response

With Caveman
User → Claude Code → Claude → Compressed response

Newer proxy functionality can go further by compressing supported input/context as well as output.

Does Caveman Reduce Claude's Intelligence?

No. Caveman primarily changes how Claude communicates, rather than replacing the underlying model.
Think:
"Let me explain the issue in detail..."
becoming:
"Root cause: DB timeout."
The goal is to remove unnecessary communication—not reduce the actual coding work.

Conclusion

For repetitive Claude Code workflows, Caveman can make the experience much cleaner:

Root cause: connection timeout.

Changed:
src/db/pool.ts

Timeout:
1s → 5s

Tests:
14/14 passed.

That's often all a developer needs.

The key takeaway is simple:

Don't assume "65% fewer output tokens" means 65% lower cost.

Test Caveman on your own workload, measure the difference, and decide whether the savings are meaningful for your workflow.

Few words. Same work. More context. Less token waste.

Comments

Post a Comment

Popular posts from this blog

LSTM and BiLSTM Explained: Advanced Deep Learning Techiniques for Time Series Prediction

Image
While RNNs are great at handling sequences, they sometimes struggle with long-term dependencies . Imagine trying to remember a detail from 20 steps ago—RNNs often “forget” that information. That’s where  LSTM (Long Short-Term Memory) and later BiLSTM (Bidirectional LSTM) models are being used. These models are game-changers for solar power forecasting. Long Short-Term Memory LSTM is an advanced type of RNN designed to remember information for longer periods. It has special structures called gates: Forget Gate: Decides what information to throw away. Input Gate: Decides what new information to store. Output Gate: Decides what the model should output at each step. LSTMs handle long-term dependencies much better than regular RNNs. Since the same dataset was used here as well, you can read more about it in my earlier RNN blog . Model Results for LSTM The training graph shows that the loss decreases steadily as the number of epochs increases. This means the model learns patterns in th...
Read more

How to Manage Secrets Securely with AWS Secrets Manager and Lambda

Introduction In modern cloud-native applications, managing sensitive information like API keys, database credentials, any third party service credentials and other secrets securely is a top priority. Hardcoding secrets into application code, environmental variables in lambda functions or configuration files can lead to serious security vulnerabilities and operational risks and this is where AWS Secrets Manager comes in—a fully managed service that enables you to store, retrieve, and rotate secrets securely. When combined with AWS Lambda , Secrets Manager allows you to build powerful serverless applications that access secrets dynamically during the runtime, without ever exposing them in your codebase. In this blog, we'll explore how to integrate AWS Secrets Manager with Lambda functions, ensuring your application remains secure, scalable, and maintainable. Whether you're accessing a database, calling a third-party service, or simply avoiding secret sprawl, this guide will wal...
Read more

Step-by-Step Guide to Setting Up AWS SES with Configuration Sets

Image
  If you're seeking a reliable and cost-effective solution for email communication then AWS Simple Email Service (SES) is an excellent choice. This cloud-based service enables you to efficiently send and receive emails with ease. Prerequisites This tutorial offers a hands-on demonstration. Ensure you have an active AWS account to follow along AWS Login . 1.  What is AWS SES AWS Simple Email Service is a cloud-based service provided by AWS that enables the business and developers to send, receive, and manage email campaigns securely and cost-efficiently. Use Cases: Transactional email notifications for applications. Marketing Campaigns Newsletters System alerts Automated Reports 2.  Why Use AWS SES? Reliable and scalable infrastructure Cost-effective compared to third-party email services Integration with other AWS services like Lambda, CloudWatch, SNS, EventBridge and Pinpoint. Built-in tools to track email events (sent, deliveries, open, click, bounces, complaints) Secur...
Read more

Using ConnectorX and DuckDB in Python: Step by Step Guide

Image
Introduction When working with large datasets, execution time and efficiency comes into play. Traditional methods of extracting data from the relational databases into Python often involve loading everything into memory, which can be painful and very slow. That’s where connectorX and DuckDB come in handy. Together, they make data extraction and analytics in python very  fast and memory-efficient . What is ConnectorX? ConnectorX is an open-source library built to load data from databases directly into pandas, Polars, or NumPy efficiently. Instead of fetching row by row via  psycopg2  or  sqlalchemy ConnectorX p arallely fetch chunks of data and stream them directly into Python. Supports many databases: MySQL, SQLite, PostgreSQL, SQL Server, BigQuery, Snowflake, and many more. What is DuckDB? DuckDB is an in-process SQL OLAP database. Can query CSV, Parquet, JSON, Arrow datasets, and even pandas/Polars DataFrames. Works directly inside Python and R. Data pro...
Read more

Kiro by AWS: The Agentic IDE That's Changing How Developers Build Software

Kiro is AWS's agentic AI IDE that goes beyond code generation — it plans, designs, executes, and documents your features as a true development partner. Built on Code OSS (the open-source foundation of VS Code), Kiro introduces spec-driven development to transform how software gets built from idea to production. This guide breaks down what Kiro is, what makes it different, and when it delivers the most value for developers. What Is Kiro? Kiro is an agentic AI IDE from AWS that understands the intent behind your prompts and orchestrates entire development workflows — planning, designing, coding, and documenting — without you having to manage any of that manually. What Kiro Handles For You Requirements documentation – Auto-generated in EARS notation from your natural language prompt System design – Analyzes your codebase and proposes architecture that fits your existing stack Task sequencing – An ordered, dependency-aware implementation plan created before a single line...
Read more

AWS Lambda: When to Use and When to Avoid Serverless Computing

Image
AWS Lambda is one of the most widely adopted serverless computing services on Amazon Web Services. It allows developers to run code without provisioning or managing servers, revolutionizing how modern applications are built and deployed. While AWS Lambda is powerful and cost-effective for many use cases, it's not suitable for every workload. This comprehensive guide will help you understand when to use AWS Lambda and when to choose alternative AWS compute services. What Is AWS Lambda? AWS Lambda is a serverless, event-driven compute service that automatically runs your code in response to events and scales instantly based on demand. What You Don't Manage Servers – No infrastructure to provision or maintain Operating systems – Automatic patching and updates Capacity planning – Automatic scaling from zero to thousands of concurrent executions Lambda Pricing Model: Pay Only for What You Use Number of invocations – First 1 million requests per month are free Execution dur...
Read more

Pandas vs Polars: Which One to Choose for Data Processing?

Image
Introduction If you’ve done any data work in Python, chances are you’ve used Pandas —it’s been the go-to library for data analysis and data preparation for years. But as datasets keep getting bigger and performance demands rise, a new player has entered the scene: Polars . Think of it as Pandas’ faster, more modern library. Both are great at handling data, but they differ quite a bit when it comes to speed, scalability, and the way they’re designed In this blog, we’ll dive into the differences between Pandas and Polars, and help you decide which one fits your use case. Pandas vs Polars Both Pandas and Polars can play an important role in data preparation and data analysis. Pandas: Pandas can integrated easily with s cikit-learn , Matplotlib, TensorFlow, and PyTorch. Built on top of  NumPy and designed for in-memory datasets Pandas is ideal for small to medium dataset. Polars: Uses  Apache Arrow memory model for efficient storage Designed to be multi-threaded and more memory...
Read more

Solar Energy Prediction Using Recurrent Neural Networks (RNN)

Image
Introduction When we think about renewable energy, solar power often comes to mind first. But there’s one big challenge: solar energy is not constant . It changes with weather, seasons, and even time of the day. To make solar power more reliable, predicting its future output becomes very important. That’s where machine learning - especially Recurrent Neural Networks (RNNs)  comes in. What is RNN A Recurrent Neural Network (RNN) is a type of deep learning model that is great at handling sequential data —data that comes in order, like time-series or speech. Unlike regular neural networks, RNNs remember past information by looping their outputs back into the network. This makes them perfect for tasks like predicting solar power output, stock prices, or even natural language processing. How RNN Works RNNs take input data step by step.  They store the output of the previous step and use it along with the new input to make better predictions.  This memory-like ability helps th...
Read more

Creating a Scalable Lambda Layer for PostgreSQL or MySQL Drivers in Python

Image
Introduction When working with AWS Lambda functions in Python, especially in database-heavy applications, you often run into deployment package size limits or performance issues due to repeated bundling of common libraries like psycopg2 for PostgreSQL, python-oracledb for Oracle or  mysql-connector-python for MySQL. These database drivers are essential, yet bulky—leading to bloated deployment packages, slower cold starts, and painful debugging across environments. To address this, Lambda Layers offer a powerful solution. Layers allow you to package shared dependencies—such as database drivers—separately and reuse them across multiple functions, simplifying deployment and improving scalability. In this blog, we’ll walk through creating a scalable and reusable Lambda Layer for PostgreSQL or MySQL drivers using Python. You’ll learn not only how to build and deploy these layers, but also best practices to make your architecture more maintainable and efficient in the long run. Whet...
Read more

Amazon S3 Files Explained (2026): How to Mount S3 Bucket as File System (Step-by-Step)

Image
AWS has introduced a major update to Amazon S3 — you can now mount your S3 bucket like a file system. This new capability, called S3 Files , changes how developers and data teams interact with storage on AWS. For years, S3 was known only as object storage. Now, it behaves much closer to a traditional file system, making it easier to work with existing tools and applications. What is Amazon S3 Files? S3 Files allows you to mount your S3 bucket as if it were a local disk on services like EC2, ECS, EKS, or Lambda. Once mounted, you can use standard file operations such as listing directories, reading files, or writing new data without any special APIs. This means developers no longer need to think in terms of objects and keys. Instead, they can simply work with files and folders, just like they would on a normal Linux system. Why This Matters Earlier, AWS storage worked like this: Amazon S3 → Cheap, scalable object storage Amazon EFS → File system for applications 👉 You had to ...
Read more