Reference File System - Token Optimization Skill

Core Principle

REFERENCE, DON'T LOAD - Treat file paths as pointers to knowledge, not content to consume.

Critical Rules

Rule 1: Path References Are Sufficient

When user mentions a file or asks to "use" something:

• DON'T automatically load the full file
• DO reference it by path: /business-memory/design-system/curv-v2.json
• ONLY load if you need to read/modify specific content

Example - CORRECT:

User: "Use the CURV design system"
Claude: "I'll reference /business-memory/design-system/curv-v2.json 
        for the design specifications."
[No file load, 20 tokens used]

Example - INCORRECT:

User: "Use the CURV design system"
Claude: [Loads entire 5KB file]
"I've reviewed your design system..."
[Full content loaded, 1,250 tokens wasted]

Rule 2: Load Only What You Need

If you must read a file:

• Use view with view_range for specific sections
• Load 10-20 lines at a time maximum
• Never load entire files "just in case"

Example - CORRECT:

view /business-memory/config.json [1-10]
# Only loads first 10 lines

Example - INCORRECT:

view /business-memory/config.json
# Loads entire 500-line file

Rule 3: File Metadata Over Content

Before loading, check file metadata:

• File exists? Use view on directory
• File size? Estimate token cost
• Can task be done without reading? Do that instead

Rule 4: Reference in Responses

When providing instructions or references:

"Apply styles from /business-memory/design-system/colors.css"
NOT: "Apply these styles: [pastes entire CSS file]"

Common Patterns

Pattern 1: Design System References

User: "Build a dashboard using CURV"
Response: "Creating dashboard with CURV design system 
          (/business-memory/design-system/curv-v2.json):
          - Primary: #9d4edd
          - Secondary: #c084fc
          - Glassmorphism effects applied"

[NO FILE LOAD - Reference system patterns from memory/previous conversations]

Pattern 2: Configuration Files

User: "Check my database config"
Response: "I'll reference your config at /business-memory/db-config.json"
[Only load if user needs specific values modified]

Pattern 3: Code Components

User: "Update the metric card component"
Response: "I'll modify the component saved at 
          /business-memory/components/metric-card.jsx"
[Load only the specific component, not entire directory]

Pattern 4: Data Files

User: "Analyze this CSV"
Response: [Load silently, process, save results]
"Analysis complete. Results saved to /mnt/user-data/outputs/analysis.xlsx"
[Don't display data unless requested]

Token Savings Calculator

File Loading Decision Tree

User mentions file
    ↓
Do I need to READ the content? 
    ↓ NO → Reference by path, continue
    ↓ YES
       ↓
Can I use cached knowledge from previous conversations?
    ↓ YES → Use cached knowledge, reference path
    ↓ NO
       ↓
Can I load just a section?
    ↓ YES → Use view_range for specific lines
    ↓ NO → Load minimum necessary, process silently

Exception Cases (When to Load)

1. User explicitly requests: "Show me what's in the file"
2. Modification required: Need to read before str_replace
3. First-time files: Never seen before, need structure understanding
4. Validation needed: Must verify file format/content
5. Complex operations: Multi-step processes requiring content awareness

Implementation Checklist

Before loading any file, ask yourself:

• Can I reference the path instead?
• Do I already know this file's content from context?
• Can I load just a section?
• Will the user actually need to see this content?
• Is this load absolutely necessary for the task?

If you answered "NO" to the last question, DON'T LOAD.

Integration with Other Skills

This skill works with:

• concise-execution-mode: Reference files, execute silently
• artifact-reference-library: Reference artifact IDs instead of recreating
• data-processing-templates: Load data silently, show results only
• incremental-file-editing: Load minimal sections for surgical edits

Success Metrics

A conversation using this skill effectively should have:

• 70-90% fewer view tool calls
• File paths mentioned frequently in responses
• Minimal file content displayed in conversation
• Fast task execution without "loading..." overhead

User Communication

When referencing files, be explicit about what you're doing:

• "I'll reference your design system at [path]"
• "Using configuration from [path]"
• "Applying patterns from [path]"

This keeps the user informed without consuming tokens with actual content.

Deployment Path: /mnt/skills/user/reference-file-system/SKILL.md Restart Required: Yes (restart Claude Desktop after deployment) Token Impact: 70-90% reduction in file-related token usage