How to Use Claude Projects to Save Tokens: Stop Repeating Context Every Session

Mar 27, 2026

Every time I started a new Claude conversation, I found myself pasting the same background information. “I’m building a React app with TypeScript, using Tailwind CSS, following these conventions…” The same 500 tokens of context, over and over, session after session. My token budget was draining on setup, not actual work. Then I discovered Claude Projects, and it changed everything.

The Problem: Token Waste Through Repetition

I was hitting Claude’s usage limits faster than expected. When I analyzed my patterns, the culprit became obvious: repetitive context. Here’s what my sessions looked like:

User: I'm building a React app with TypeScript, using Tailwind CSS,
following these conventions... [500 tokens of context]... now help me with X

User: I'm building a React app with TypeScript, using Tailwind CSS,
following these conventions... [500 tokens of context]... now help me with Y

User: I'm building a React app with TypeScript, using Tailwind CSS,
following these conventions... [500 tokens of context]... now help me with Z

Each session, I burned through tokens just setting up context. Claude had no memory of my previous conversations, so I had to re-explain my project every time. This wasn’t just annoying - it was expensive.

The math was brutal:

Average context paste: 500-1,000 tokens
Sessions per day: 5-10
Daily waste: 2,500-10,000 tokens
Monthly waste: 75,000-300,000 tokens

That’s a significant portion of my token budget spent on repetition.

The Solution: Claude Projects for Persistent Context

I learned about Claude Projects from a Reddit discussion titled “10 Tricks to Stop Hitting Claude’s Usage Limits.” Trick #2 mentioned:

“If you’re repeatedly pasting the same background info (‘I’m a Python dev, my codebase uses X, my tone is Y’), put it in a Project system prompt. Stop wasting tokens re-explaining yourself every session.”

This was the game changer I needed.

What Claude Projects Actually Does

Claude Projects creates a persistent workspace with three key features:

Project-level system prompt: Instructions that apply to ALL conversations in that project
Project knowledge base: Documents and files Claude can reference
Conversation history: All chats within the project are searchable

The magic is in feature #1. Instead of pasting context in every conversation, you add it ONCE to the Project’s system prompt. This context is automatically included without consuming tokens from your conversation budget.

Setting Up My First Project

I went to Claude.ai and clicked “Projects” in the sidebar. Here’s the setup process I followed:

Step 1: Create a new Project

I named it “E-commerce Platform” for my headless e-commerce project.

Step 2: Add a custom instruction (system prompt)

This is where the token savings happen. I added my persistent context:

## Project Context
Building a headless e-commerce platform with multi-vendor support.

## Architecture
- Frontend: Next.js 14 App Router, TypeScript
- Backend: NestJS, GraphQL API
- Database: MongoDB with Mongoose
- Payments: Stripe Connect

## Key Files (when I reference them)
- `/lib/stripe-helpers.ts` - Stripe utility functions
- `/components/checkout/` - Checkout flow components
- `/pages/api/webhooks/` - Stripe webhook handlers

## My Preferences
- Use React Query for data fetching
- Implement optimistic updates
- Add loading states for all async operations
- Follow the existing component structure

## Coding Standards
- Functional components with hooks
- Zod for validation
- Repository pattern for data access
- Tests required for all new features

When helping me:
- Suggest production-ready code, not tutorials
- Include error handling
- Follow existing patterns in the codebase
- Flag security concerns immediately

Step 3: Start all related conversations within this Project

Now when I need help, I just ask directly without any setup.

The Results: Massive Token Savings

The difference was immediate. Here’s a comparison:

// User pastes: "I'm using Next.js 14 with App Router, TypeScript,
// Tailwind, and Prisma. My database is PostgreSQL. I need server
// components by default, client components only for interactivity.
// Follow these coding standards... [continues for 1000 tokens]"
//
// Then asks: "Help me add a search feature"
//
// Total tokens: ~1,500 for setup + question

// User just asks: "Help me add a search feature"
// Project context automatically included
//
// Total tokens: ~100 for the question
// Savings: ~1,400 tokens per conversation

My actual savings over a week:

Before Projects: ~75,000 tokens used (mostly setup repetition)
After Projects: ~15,000 tokens used (pure productive work)
Savings: ~60,000 tokens in one week

I stopped hitting usage limits. My conversations became more focused. Claude remembered my preferences every session.

Common Mistakes I Made (So You Don’t Have To)

Mistake 1: Over-stuffing the System Prompt

My first attempt included everything about my project - every API endpoint, every database schema, every coding standard. The system prompt was 3,000 tokens long. This made Claude slower and sometimes confused its responses.

Lesson learned: Keep the system prompt under 2,000 tokens. Focus on what Claude needs to know to help you, not everything about your project.

Mistake 2: Not Organizing by Project

I initially created one giant Project for all my work. This led to context conflicts - Claude would suggest Python solutions for my TypeScript project.

Lesson learned: Create separate Projects for different codebases/clients. One Project per major initiative. Keep personal and work projects separate.

Mistake 3: Forgetting to Update Context

When I switched from REST to GraphQL, I forgot to update my Project’s system prompt. Claude kept suggesting REST patterns.

Lesson learned: When your stack changes, update the system prompt immediately. Remove outdated information to keep responses relevant.

Mistake 4: Not Using Project Documents

I didn’t upload any reference documents to my Projects. I kept pasting API specs and coding standards in conversations.

Lesson learned: Upload reference docs, API specs, and coding standards to the Project. Claude can search these when answering questions, reducing token usage even further.

Beyond Token Savings: Additional Benefits

While token savings were my primary motivation, I discovered other benefits:

Consistency: Claude remembers my preferences every session. No more explaining “I prefer functional components” for the tenth time.

Efficiency: I start working immediately. No setup required. The flow is: open Project, ask question, get help.

Context accumulation: I add to the project knowledge over time. As my project grows, so does Claude’s understanding of it.

Better responses: Claude has full project context for every question. The suggestions are more relevant and aligned with my architecture.

When to Create a New Project

I now follow these rules for when to create a new Claude Project:

Different codebase - Each major project gets its own Project
Different client - Client work is always in separate Projects
Different tech stack - If the stack changes significantly, new Project
Long-term project - If I’ll be working on it for weeks, it deserves a Project

If I’m just asking a one-off question or exploring an idea, I use a regular conversation. But as soon as I find myself pasting context twice, I create a Project.

Summary

In this post, I shared how Claude Projects transformed my workflow from token-wasting repetition to efficient productivity. The setup takes 5 minutes, but the savings are immediate and substantial. By storing persistent context in a Project’s system prompt, I save thousands of tokens per month, get consistent responses, and stay under usage limits.

The key insight: Projects = Persistent Context = Token Efficiency. Stop burning tokens on repetition and start building on your accumulated knowledge.

If you find yourself pasting the same context more than twice, create a Project for it. Your token budget (and sanity) will thank you.

Final Words + More Resources

My intention with this article was to help others share my knowledge and experience. If you want to contact me, you can contact by email: Email me

Here are also the most important links from this article along with some further resources that will help you in this scope:

Oh, and if you found these resources useful, don’t forget to support me by starring the repo on GitHub!