long-context-processing

Long Context Processing with Gemini

When to Use This Skill

Automatically invoke this skill when:

• User needs to analyze an entire codebase at once
• Document or file content exceeds ~100K tokens
• User explicitly requests processing multiple large files together
• Task requires comprehensive "see everything" context
• User asks to find patterns across many files
• Analyzing very long documents (books, research papers, extensive logs)

Examples That Trigger This Skill

• "Analyze the entire codebase architecture"
• "Summarize this 500-page document"
• "Find patterns across all log files"
• "Review all Python files in this project"
• "What are the common themes in these research papers?"
• "Analyze all API endpoints across the codebase"

How to Use

1. Gather content:
   • Use Glob to find all relevant files
   • Use Read to get file contents
   • Concatenate into a single large text block
2. Estimate size: Calculate approximate tokens (chars / 4)
3. Call Gemini: Use the process_long_context tool from gemini-api MCP server
   • Pass the aggregated content
   • Include user's analysis request as the prompt
   • Use gemini-1.5-pro for best long-context performance
4. Present results: Return Gemini's comprehensive analysis

Tool Parameters

{
  "content": "[very large concatenated text content]",
  "prompt": "Analyze the overall architecture and identify main patterns",
  "model": "gemini-1.5-pro"  // Pro recommended for long context
}

Capabilities

• Context Window: Up to 2 million tokens (~1.5 million words, ~8 million characters)
• Whole Codebase Analysis: See entire project structure and relationships
• Pattern Detection: Find recurring themes across extensive content
• Comprehensive Summarization: Distill key points from massive documents
• Cross-file Analysis: Understand how different parts relate to each other

When NOT to Use

• Small documents (< 50K tokens) - use Claude directly instead
• When detailed code editing is needed - Claude is better for precise changes
• Real-time or interactive tasks - prefer Claude's faster response
• When summarization is not the primary goal

Best Practices

• Always use gemini-1.5-pro model for long context tasks
• Structure the prompt clearly: "Given the following [codebase/document/logs], [task]"
• Include file paths or section markers in content for better reference
• Ask for structured output (headings, lists) for easier parsing
• For very large content, consider if you truly need ALL of it or can filter first

Content Formatting

When aggregating multiple files, use clear delimiters:

=== FILE: path/to/file1.py ===
[file content]

=== FILE: path/to/file2.py ===
[file content]

This helps Gemini maintain file context in its analysis.