Context Management for AI Coding

Techniques for managing LLM context windows effectively when working with AI coding assistants.

Overview

Context management is the art of providing LLM coding assistants with the right information at the right time. Models have limited context windows, so strategic inclusion and exclusion of code, docs, and instructions directly impacts output quality.

Context Window Sizes

Model	Context Window	Approx. Lines of Code
Claude 3.5 Sonnet	200K tokens	~50,000 lines
GPT-4 Turbo	128K tokens	~32,000 lines
Gemini 1.5 Pro	1M tokens	~250,000 lines
Llama 3.1	128K tokens	~32,000 lines

Strategies

•Provide only the relevant files — not the entire codebase
•Summarize large modules with their public API signatures
•Use file headers to describe purpose and dependencies
•Front-load the most important context (instructions, types)
•Remove conversation history that is no longer relevant
•Split large tasks into smaller, focused conversations

File Selection Guide

Bug Fix

Include: the buggy code, its tests, related type definitions, and the error message.

New Feature

Include: the nearest similar feature, relevant types, and the routing/API layer where it integrates.

Refactoring

Include: the code to refactor, its tests, and any code that imports/depends on it.

Frequently Asked Questions

Why does context management matter?

LLMs have limited context windows. Providing the right code at the right time is critical — too much irrelevant context dilutes quality, too little causes hallucination.

How large are current LLM context windows?

Claude supports 200K tokens, GPT-4 supports 128K, and Gemini supports 1M+. Even with large windows, focused context produces better results.