Context Management Intermediate
Context is king in AI-assisted development. Understanding how Claude Code reads your codebase, managing conversation length, and providing the right context at the right time is what separates good results from great ones.
How Claude Code Reads Your Codebase
Claude Code does not read your entire codebase at once. It uses tools to explore on-demand:
-
CLAUDE.md is read first
At session start, Claude reads your
CLAUDE.mdfile for project context, conventions, and instructions. -
Tools are used to explore
Claude uses
Read,Grep,Glob, andBashtools to find and read relevant files based on your prompt. -
Context accumulates in the conversation
Every file read, every command output, and every response adds to the conversation context window.
-
Context window has limits
When the conversation gets too long, older context may be truncated. Use
/compactto summarize and free up space.
CLAUDE.md: The Ultimate Context File
Your CLAUDE.md file is the single most important factor in Claude Code's effectiveness. It is read at the start of every session and provides persistent context.
# Project: E-Commerce API ## Tech Stack - Node.js 20 + TypeScript 5.3 - Fastify (NOT Express) - PostgreSQL 16 + Prisma ORM - Jest for testing - pnpm (NOT npm or yarn) ## Key Commands - `pnpm dev` - Start dev server - `pnpm test` - Run all tests - `pnpm test:watch` - Watch mode - `pnpm db:migrate` - Run migrations - `pnpm lint` - ESLint + Prettier ## Conventions - Use barrel exports (index.ts) in each directory - Error handling: throw custom errors from src/errors/ - All dates in ISO 8601 format, stored as UTC - API responses follow the format in src/types/api-response.ts ## File Structure - src/routes/ - API route handlers - src/services/ - Business logic - src/repos/ - Database queries (Prisma) - src/middleware/ - Fastify middleware - src/types/ - TypeScript type definitions
Using /compact to Manage Long Sessions
As your conversation grows, Claude's context window fills up. The /compact command summarizes the conversation and frees up space:
# When you notice Claude forgetting earlier context, or after many turns > /compact # Claude summarizes the conversation and frees up context space # You can continue the session with the summary as context # Pro tip: use /compact with a custom summary prompt > /compact Focus on the database migration work we discussed
When to Start a New Session
Sometimes it is better to start fresh rather than compact:
| Start New Session When... | Use /compact When... |
|---|---|
| Switching to a completely different task | Continuing the same task but running low on context |
| Claude is confused or making repeated errors | You've been working on a long refactoring |
| You want a fresh perspective on a problem | You need to keep the decisions made so far |
| The conversation is about a topic unrelated to previous context | Claude's responses are getting slower (context is large) |
Providing File Paths Explicitly
Every time Claude uses a tool to search for files, it costs tokens and time. Providing file paths saves both:
# Expensive: Claude must search for the file > Fix the email validation bug # Cheap: Claude goes directly to the file > Fix the email regex in src/utils/validators.ts line 23 # Even better: reference multiple files explicitly > Update the User interface in src/types/user.ts to add an avatar field, > then update src/services/user.service.ts and src/routes/users.ts to handle it
Using Git History as Context
Git history provides rich context about your project's evolution:
# Use recent commits as context > Look at the last 5 commits. The feature we added yesterday is causing a regression. > Find and fix it. # Use blame for understanding code authorship > Check git blame on src/services/payment.ts. The logic added in the most recent > commit is incorrect. Revert just that function to the previous version. # Use diff for targeted reviews > Review git diff main...feature-branch. Focus on the database queries.
Reducing Noise: What NOT to Include
Too much context is almost as bad as too little. Avoid these context polluters:
- Generated files: Don't let Claude read
node_modules/,dist/, orbuild/directories. - Lock files:
package-lock.jsonandyarn.lockare huge and rarely useful. - Binary files: Images, fonts, and compiled assets waste context space.
- Verbose logs: Pipe only the relevant portion of logs, not entire files.
- Irrelevant history: Don't dump the entire git log when Claude only needs the last few commits.
The Context Budget Mental Model
Think of your context window as a budget. Every interaction has a cost:
| Action | Context Cost | Optimization |
|---|---|---|
| Reading a file | ~1 token per character | Reference specific line ranges when possible |
| Running a command | Varies by output size | Pipe output through head or grep |
| Your prompt | ~1 token per 4 characters | Be concise but specific |
| Claude's response | ~1 token per 4 characters | Ask for brief responses when appropriate |
| CLAUDE.md | Read once per session | Keep it focused (under 500 lines) |
Memory System Usage
Claude Code has a memory system that persists information across sessions:
# Claude automatically remembers project-level preferences # You can also explicitly tell Claude to remember things > Remember: this project uses pnpm, not npm. Always use pnpm commands. > Remember: our API responses always include a "success" boolean field. # These memories persist in CLAUDE.md or Claude's memory system # and are available in future sessions
Try It Yourself
Create or update your project's CLAUDE.md with tech stack, key commands, and conventions. Then start a Claude Code session and notice how it immediately understands your project better. Next up: speed and efficiency tips.
Ready to Go Deeper?
Live instructor-led courses from our partners. Affiliate disclosure.
AI & ML Courses - 30% Off
Live instructor-led AI, machine learning, data science, and cloud courses for working professionals. Use code Limited30 at checkout.
EdurekaDataCamp - AI & Data Science
Hands-on Python, machine learning, and AI courses with interactive exercises and real projects.
DataCampedX - Top AI Courses
University-level AI courses from MIT, Harvard, Stanford. Earn certificates that employers recognize.
edX