Look At - Multimodal File Analysis

Look At - Multimodal File Analysis

Fast, cost-effective file analysis using Google's Gemini 2.5 Flash Lite model for PDFs, images, diagrams, and other media files.

Tool Selection Enforcement

Rationalization Table - STOP When Thinking:

Red Flags - STOP Immediately When Thinking:

• If you catch yourself thinking "Let me Read this image/​PDF/​screenshot" → STOP. Use look_at for media files.
• If you catch yourself thinking "I can see the image directly" → STOP. Seeing it directly still wastes context. Use look_at.
• If you catch yourself thinking "Just need to glance at this diagram" → STOP. Glancing still costs context tokens. Use look_at.
• If you catch yourself thinking "The PDF is text-based, so Read is fine" → STOP. If it has structure/​tables/​charts, use look_at.

Cost & Context Benefits

look_at saves 80-95% of context tokens by extracting only relevant information.

When to Use

Use look_at when you need:

• Media files the Read tool cannot interpret
• Extracting specific information or summaries from documents
• Describing visual content in images or diagrams
• Analyzing charts, tables, or structured data in PDFs
• When analyzed/​extracted data is needed, not raw file contents

Never use look_at when:

• Source code or plain text files needing exact contents (use Read)
• Files that need editing afterward (need literal content from Read)
• Simple file reading where no interpretation is needed
• Exact formatting or structure must be preserved

How It Works

1. Provide a file path and a specific goal (what to extract)
2. The helper script uploads the file to Gemini's API
3. Gemini 2.5 Flash Lite analyzes the file and extracts requested information
4. Only the relevant extracted information is returned (saves context tokens)

Usage Pattern

CRITICAL - Display Requirement: Always set the Bash tool description parameter to show a clean invocation:

description: "look-at: [goal text]"

Never display the full Python command to the user.

# Basic usage
python3 ${CLAUDE_PLUGIN_ROOT}/​skills/​look-at/​scripts/​look_at.py \
    --file "/​path/​to/​file.pdf" \
    --goal "Extract the title and date from this document"

# With custom model
python3 ${CLAUDE_PLUGIN_ROOT}/​skills/​look-at/​scripts/​look_at.py \
    --file "/​path/​to/​diagram.png" \
    --goal "Describe the architecture shown in this diagram" \
    --model "gemini-2.5-flash"

IMPORTANT:

• Always use absolute paths for files
• Always set Bash tool description to "look-at: [goal]" for clean UX

Response Rules

When using look_at, the response includes:

• Only the extracted information matching the goal
• Clear statement if requested information is not found
• Concise output focused on the goal (no preamble)

Use this extracted information directly in continued work without loading the full file into context.

Supported File Types

Model Options

Default is gemini-2.5-flash-lite for optimal speed/​cost ratio.

Common Patterns

REMEMBER: Always use description: "look-at: [goal]" in the Bash tool call.

Extract Specific Information

# Bash tool call with:
# description: "look-at: Extract the executive summary section"
python3 ${CLAUDE_PLUGIN_ROOT}/​skills/​look-at/​scripts/​look_at.py \
    --file "report.pdf" \
    --goal "Extract the executive summary section"

Describe Visual Content

# Bash tool call with:
# description: "look-at: List all UI elements and their layout"
python3 ${CLAUDE_PLUGIN_ROOT}/​skills/​look-at/​scripts/​look_at.py \
    --file "screenshot.png" \
    --goal "List all UI elements and their layout"

Analyze Diagrams

# Bash tool call with:
# description: "look-at: Explain the data flow and component relationships"
python3 ${CLAUDE_PLUGIN_ROOT}/​skills/​look-at/​scripts/​look_at.py \
    --file "architecture.png" \
    --goal "Explain the data flow and component relationships"

Extract Structured Data

# Bash tool call with:
# description: "look-at: Extract the table data as JSON"
python3 ${CLAUDE_PLUGIN_ROOT}/​skills/​look-at/​scripts/​look_at.py \
    --file "table.pdf" \
    --goal "Extract the table data as JSON with columns: name, value, date"

Environment Setup

Required environment variable:

export GOOGLE_API_KEY="your-api-key-here"

Required Python package:

pip install google-genai

For pixi-managed projects, add to pixi.toml:

[dependencies]
google-genai = ">=1.0.0"

Cost Optimization

• Gemini 2.5 Flash Lite is the most cost-effective option
• Only extracts requested information (saves on output tokens)
• Avoids loading full files into main conversation context
• Use specific goals to minimize unnecessary processing

Troubleshooting

Examples

See examples/ directory for:

• analyze_pdf.sh - PDF document extraction
• describe_image.sh - Image analysis
• extract_table.sh - Structured data extraction

Related Skills

• /​gemini-batch - For batch processing of many files
• Standard Read tool - For text files needing exact contents