Understanding Context

Context is the foundation of how Large Language Models (LLMs) work and directly affects the quality and relevance of the Gamecraft Agent's responses.Context is the foundation of how Large Language Mod

This guide explains how context works, how to interpret the context information shown in the UI, and how to optimize context for better performance.

What is Context?

Context in Large Language Models

Context refers to all the information that the AI model can "see" and remember during a conversation. Think of it as the AI's working memory - everything it knows about your current task, conversation history, and project state.

Unlike humans who can remember things indefinitely, LLMs have a fixed context window - a limit to how much information they can process at once. This is measured in tokens, which are roughly equivalent to words (though technical terms and code may use more tokens).

Context Components in Gamecraft Agent

When you interact with the Gamecraft Agent, your context includes:

  1. System Prompt: Instructions that define the agent's personality and capabilities

  2. Conversation History: Your messages and the agent's responses

  3. Project Information: File contents, directory structure, terminal output

  4. Environment Details: Operating system, tools available, workspace settings

  5. Task Context: Current mode, previous actions, and ongoing objectives

Context Window Visualization

Understanding the Progress Bar

The Gamecraft Agent displays your context usage through a visual progress bar in the task header:

Context Window: [████████████▓▓▓░░░] 12.5K / 128K tokens
                ^Current    ^Reserved ^Available

Color-Coded Sections

  • Dark Gray: Tokens currently used by your conversation and project data

  • Medium Gray: Tokens reserved for the model's response

  • Light Gray: Available space for additional context

Hover Information

Hover over different sections of the progress bar to see:

  • Current Section: "12,500 tokens used of 128,000 total"

  • Reserved Section: "Reserved for model response: 25,600 tokens"

  • Available Section: "Available space: 89,900 tokens"

Token Count Display

Below the progress bar, you'll see detailed token metrics:

  • ↑ Input Tokens: What you've sent to the AI (your messages + project context)

  • ↓ Output Tokens: What the AI has generated in responses

  • Cache Writes/Reads: For models supporting prompt caching (faster, cheaper processing)

Context Management Strategies

Auto-Context Condensing

When your context window fills up, Gamecraft Agent can automatically manage it:

How It Works

  1. Threshold Detection: When context usage exceeds your configured percentage (default: 100%)

  2. Intelligent Summarization: The AI creates a concise summary of older conversation parts

  3. Context Replacement: Older messages are replaced with the summary, freeing up space

  4. Continuation: Your task continues with full context awareness

Configuration Options

In Settings → Context Management:

  • Auto Condense Context: Enable/disable automatic condensing

  • Trigger Percentage: When to start condensing (50-100% of context window)

  • Condensing API: Which AI model to use for summarization

  • Custom Prompt: Personalize how summaries are created

Manual Context Management

Condense Context Button

Click the Fold button (🗜️) in the task header to manually trigger context condensing:

  • Useful before starting complex operations

  • Helps prevent mid-task interruptions

  • Preserves all important information while reducing token usage

Starting Fresh

When context becomes unwieldy:

  1. Export Current Task: Save your progress before starting over

  2. Close Task: End the current conversation

  3. New Task: Begin with clean context, referencing exported work if needed

Context Optimization Techniques

Effective Context Usage

Be Selective with File Mentions

Instead of including entire directories:

# Less Efficient
@src/ @tests/ @docs/
# More Efficient  
@src/components/UserAuth.tsx @tests/auth.test.ts

Use Specific References

Target exactly what you need:

# Vague
"Look at my authentication system"
# Specific
"Review @src/auth/loginHandler.ts for the JWT validation logic"

Strategic Information Sequencing

Start with the most important context:

  1. Core Problem: What you're trying to solve

  2. Relevant Files: Specific code that needs attention

  3. Error Messages: If debugging

  4. Additional Context: Background information if space allows

Context-Aware Prompt Writing

Front-Load Important Information

# Good Structure
Fix the authentication bug in @src/auth.ts where users can't log in 
with special characters in passwords. The error occurs in the 
validatePassword function around line 45.

[Additional details if context allows...]

Break Down Complex Tasks

Instead of one massive prompt:

# Large Task (may exceed context)
"Refactor the entire user management system to use microservices, 
update all authentication flows, migrate the database schema, 
update tests, and deploy to staging"

Break it into focused steps:

# Step 1
"Plan the microservices architecture for user management. 
Identify the core services needed and their responsibilities."

# Step 2 (new task after completing step 1)
"Implement the user authentication service based on the 
architecture we planned."

Context Limits and Performance

Model Context Windows

Different AI models have varying context capacities:

Model
Typical Context Window

GPT-4

32K - 128K tokens

GPT-4o

128K tokens

Claude 3.5 Sonnet

200K tokens

Gemini 2.5 Pro

1M - 2M tokens

Local Models

4K - 32K tokens

Performance Considerations

Larger Context ≠ Always Better

  • Cost: More tokens = higher API costs

  • Speed: Larger contexts take longer to process

  • Focus: Too much information can dilute the AI's attention

  • Quality: Relevant context beats comprehensive context

Sweet Spot Guidelines

For optimal performance:

  • Keep conversations focused on specific problems or features

  • Use 50-70% of available context for best balance of information and performance

  • Condense regularly for long-running tasks

  • Start fresh for significantly different problems

Troubleshooting Context Issues

Common Problems

"Context Window Full" Errors

Symptoms: Task stops mid-conversation, context condensing messages Solutions:

  • Enable auto-condensing with lower threshold (70-80%)

  • Manually condense before starting complex operations

  • Start a new task for unrelated work

Poor Response Quality

Symptoms: Agent seems confused, references irrelevant information Solutions:

  • Check if important context was condensed away

  • Provide fresh context for the current problem

  • Use more specific file references

Slow Response Times

Symptoms: Long delays before agent responds Solutions:

  • Reduce context by condensing or starting fresh

  • Use smaller, more focused prompts

  • Consider switching to a model with larger context window

High API Costs

Symptoms: Unexpected billing from AI provider Solutions:

  • Enable aggressive auto-condensing (50-60% threshold)

  • Break large tasks into smaller, focused sessions

  • Use models with lower per-token costs for routine tasks

Optimization Settings

In Settings → Context Management, configure:

File Context Limits

  • Max Open Tabs: Limit files automatically included (default: 20)

  • Max Workspace Files: Limit files in project context (default: 200)

  • Max File Lines: Truncate very large files

  • Concurrent File Reads: Balance speed vs. resource usage

Auto-Condensing

  • Enable Auto Condense: Automatic context management

  • Condense Threshold: When to trigger (recommend 70-80%)

  • Condensing Model: Separate API for cost optimization

  • Custom Condensing Prompt: Tailor summarization style

Best Practices Summary

For New Users

  1. Monitor the context bar - Watch how your usage grows

  2. Enable auto-condensing at 80% threshold to start

  3. Use specific file mentions rather than broad directory includes

  4. Start new tasks for different features or problems

For Power Users

  1. Customize condensing prompts for your workflow

  2. Use different models for condensing vs. main tasks

  3. Export task history before major context changes

  4. Batch similar operations to maximize context efficiency

For Teams

  1. Establish context guidelines for shared projects

  2. Document important decisions outside of context-dependent conversations

  3. Use consistent file organization to improve context relevance

  4. Share effective prompt patterns that work well with your codebase

Context and Output Quality

How Context Affects Agent Performance

Relevant Context = Better Results

The agent performs best when context includes:

  • Directly related code to the task at hand

  • Recent conversation history about the current problem

  • Error messages and debugging information

  • Project structure and conventions

Information Overload

Too much context can lead to:

  • Confused responses mixing unrelated concepts

  • Generic solutions instead of project-specific ones

  • Missed details as important information gets lost

  • Higher costs without proportional quality gains

Context Quality Over Quantity

Focus on including:

  1. Precise problem descriptions

  2. Relevant code sections with proper context

  3. Expected behavior vs. actual behavior

  4. Previous attempted solutions to avoid repetition

Understanding and optimizing context usage will significantly improve your experience with the Gamecraft Agent, leading to more accurate responses, lower costs, and faster development cycles.

Next Steps

Now that you understand the context window, feel free to check out more resources for optimizing prompts and agent context

  1. Prompt Engineering Guide - Targeted guide for improving prompts passed to Gamecraft Agent

  2. Avoid Context Poisoning - Guidance to keep the agent's context tied to reality

PreviousYour First TaskNextPrompt Engineering Guide

Last updated