Look At - Multimodal File Analysis

What this skill carries. The names and headings are the index; for a subject none of them carries, grep -il <term> over this skill's own references/ directory.

!d=${CLAUDE_SKILL_DIR}; command -v skill-toc >/dev/null 2>&1 && exec skill-toc "$d"; s=$HOME/.claude/skills/plugin-utils/bin/skill-toc; [ -x "$s" ] && exec "$s" "$d"; echo "(skill-toc unavailable: references and scripts are NOT listed here — install the plugin-utils plugin, or start a new session so its bin/ reaches PATH)"

Multi-backend vision router for images, PDFs, video, diagrams and other media. Defaults to agy -p on the vision_antigravity role in ${CLAUDE_PLUGIN_ROOT}/scripts/lib/gemini-models.json — Gemini via Antigravity OAuth, unmetered — which reads images, PDFs and video natively. Audio auto-routes to the metered api backend, the only one that handles it. Three further unmetered CLI backends (claude-code -p, codex exec, Copilot on GPT-5.4) give independent second opinions.

Tool Selection Enforcement

Tool Routing Facts

• Read on a media file loads the full content into context regardless of how briefly you look at it — a "quick glance" costs the same thousands of tokens as a full read. Content type, not file size, determines the tool.
• Read on a PDF extracts raw text and loses table structure and visual information; look_at returns it as structured data.
• The point is context economy, not vision capability. Read pulls the whole image into this session's context; look_at spends a subprocess's context instead and returns text. Better vision models do not change that arithmetic — they make the cheap backends sufficient.
• Backend extraction is accurate for most use cases — start with look_at, escalate to Read only if the extraction is insufficient. Defaulting to Read "for exact text" wastes the context this skill exists to save.
• The claude backend spawns a child claude-code -p. look_at.sh sets LOOK_AT_NESTED=1 so image-read-guard.ts stands down inside that child — without it the guard denies the child's Read and points it back at look_at.sh, which spawns another child. That is unbounded recursion, not a slow call.

Red Flags

• Passing an image, PDF, or screenshot path to Read → use look_at.
• A text-based PDF with structure/tables/charts → still look_at, not Read.

Cost & Context Benefits

Scenario	Read Tool	look_at Tool
PDF with table	Extracts raw text (~1000 tokens), loses table structure	Extracts table as structured data (~100 tokens)
Screenshot	Loads entire image (~500 tokens), requires interpretation	Describes content (~50 tokens)
Diagram	Shows image (~800 tokens), requires analysis	Explains architecture (~100 tokens)
Multi-page PDF	All pages loaded (~5000 tokens)	Extracts specific sections (~200 tokens)

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. look_at.sh routes to the selected backend (agy by default)
3. The backend 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]"

# Default (agy — Gemini via Antigravity OAuth, unmetered)
"${CLAUDE_SKILL_DIR}/scripts/look_at.sh" \
    --file "/path/to/file.pdf" \
    --goal "Extract the title and date from this document"

# A different model family (agy, codex, copilot all work the same way)
"${CLAUDE_SKILL_DIR}/scripts/look_at.sh" \
    --file "/path/to/diagram.png" \
    --goal "Describe the architecture" \
    --backend codex

# Four independent looks at once (claude, agy, codex, copilot)
"${CLAUDE_SKILL_DIR}/scripts/look_at.sh" \
    --file "/path/to/diagram.png" \
    --goal "Score this diagram 0-10" \
    --consensus

# PDFs and video need no flags — agy reads both natively
"${CLAUDE_SKILL_DIR}/scripts/look_at.sh" \
    --file "/path/to/file.pdf" \
    --goal "Extract the table data"

# Agentic mode — adds code execution for harder visual reasoning (api backend only)
"${CLAUDE_SKILL_DIR}/scripts/look_at.sh" \
    --file "/path/to/file.pdf" \
    --goal "Extract the table data" \
    --agentic

${CLAUDE_SKILL_DIR} is substituted at skill load time, so the full path is already resolved — no per-call discovery needed.

IMPORTANT:

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

Backends

Backend	CLI	Model	Cost	Best For
claude	claude-code -p	claude-opus-5[1m] unless --model	Pooled OAuth via CLIProxyAPI	Unmetered second opinion from a different family
agy (default)	agy -p	role vision_antigravity unless --model	Antigravity OAuth — unmetered	Images, PDFs and video, all read natively. No audio
codex	codex exec	Codex default	Subscription	Attaches the image with -i, so it needs no read tool at all
copilot	copilot -p	GPT-5.4	Copilot subscription	Fourth opinion. PDFs rasterized first
api	look_at.py	role vision, thinking_level=high	Metered — your GOOGLE_API_KEY	Audio auto-routes here — no unmetered backend handles it. Not in --consensus

claude, not claude-code, is the backend name; claude-code is the binary it runs. Plain claude would bill this session's own account — claude-code routes through CLIProxyAPI to the pooled OAuth accounts, which is the cost this backend exists to avoid.

Only claude ingests PDFs directly. agy and copilot get page PNGs from pdftoppm; codex gets every page attached as a separate -i.

The old gemini backend is gone — it billed like api despite being documented as bundled quota, and the consumer gemini binary was sunset 2026-06-18. Unmetered Gemini now comes from agy (Antigravity OAuth), which is the default. gemini-code is NOT usable here: on Claude Code 2.1.238 it exits 0 having produced no output, rejecting its own default model as unrecognized_model even though the proxy serves it.

Consensus Mode

--consensus runs a comma-separated list of backends in parallel and outputs each result under a labeled header (=== CLAUDE (claude-code) ===, === AGY (Antigravity) ===, …). The list is optional and defaults to all four CLI backends:

--consensus                            # claude,agy,codex,copilot
--consensus claude,codex               # narrow it to two

Wall-clock is the slowest backend, not the sum — they run concurrently.

A failed backend prints [ERROR] <name> backend failed followed by its output; the others still report, and the exit status stays 0.

When to use: Visual verification of diagrams where a single model may miss or underscore defects. Trust the stricter score — if any backend flags BLOCKING, treat it as BLOCKING.

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

Type	Extensions	MIME Types
Images	.jpg, .jpeg, .png, .webp, .heic, .heif	image/*
Videos	.mp4, .mpeg, .mov, .avi, .webm	video/*
Audio	.wav, .mp3, .aiff, .aac, .ogg, .flac	audio/*
Documents	.pdf, .txt, .csv, .md, .html	application/pdf, text/*

Model Options (api backend only)

These apply to --backend api, which is metered. The claude backend takes --model as a Claude model alias; copilot is pinned to GPT-5.4.

Model	Use Case	Speed	Cost
gemini-3.7-flash	Default - most capable stable Flash, thinking_level=high	Fast	$0.75/1M
gemini-3.5-flash-lite	High-throughput / document parsing when cost matters	Fastest	$0.30/1M
gemini-3.1-pro-preview	Maximum vision capability, hardest extractions	Slower	$2.00/1M
gemini-3-pro-preview	Highest accuracy required	Medium	Medium

The default is the vision role in ${CLAUDE_PLUGIN_ROOT}/scripts/lib/gemini-models.json, at thinking_level=high; the agy backend uses vision_antigravity, whose ids carry a reasoning suffix and are not interchangeable with the API ids above. python3 ${CLAUDE_PLUGIN_ROOT}/scripts/lib/gemini_models.py <role> prints either.

Agentic Vision Mode (api backend only)

For complex visual reasoning tasks, use the --agentic flag to enable code execution. This allows Gemini to:

• Zoom into specific regions of an image for detailed analysis
• Count objects precisely using programmatic analysis
• Perform calculations on visual data (measurements, statistics)
• Process structured data in images (charts, tables) with higher accuracy

When to use --agentic:

• Counting objects in an image ("How many items are in this photo?")
• Reading fine details ("What does the small text in the corner say?")
• Analyzing charts with specific data points ("What's the exact value for Q3?")
• Complex spatial reasoning ("Which element is closest to the center?")

Usage:

"${CLAUDE_SKILL_DIR}/scripts/look_at.sh" \
    --file "photo.jpg" \
    --goal "Count the number of people in this image" \
    --agentic

Note: Agentic mode adds code execution to whichever model is selected; gemini-3.7-flash supports it, so it no longer forces a different model.

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"
"${CLAUDE_SKILL_DIR}/scripts/look_at.sh" \
    --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"
"${CLAUDE_SKILL_DIR}/scripts/look_at.sh" \
    --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"
"${CLAUDE_SKILL_DIR}/scripts/look_at.sh" \
    --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"
"${CLAUDE_SKILL_DIR}/scripts/look_at.sh" \
    --file "table.pdf" \
    --goal "Extract the table data as JSON with columns: name, value, date"

Count Objects (Agentic)

# Bash tool call with:
# description: "look-at: Count the number of people in the photo"
"${CLAUDE_SKILL_DIR}/scripts/look_at.sh" \
    --file "crowd.jpg" \
    --goal "Count the number of people visible in this image" \
    --agentic

Analyze Chart Details (Agentic)

# Bash tool call with:
# description: "look-at: Extract specific data points from the chart"
"${CLAUDE_SKILL_DIR}/scripts/look_at.sh" \
    --file "quarterly_chart.png" \
    --goal "Extract the exact values for each quarter and calculate the year-over-year change" \
    --agentic

Environment Setup

The four CLI backends need nothing beyond their own binaries being installed and signed in (claude-code, agy, codex, copilot). No API key, no Python environment. agy, codex and copilot additionally need pdftoppm (poppler-utils) to accept a PDF.

Only --backend api needs setup, and only because it is metered:

export GOOGLE_API_KEY="your-api-key-here"   # or GEMINI_API_KEY

look_at.sh launches it with uv run --script, which honours look_at.py's inline PEP 723 metadata and provisions google-genai itself — uv run python3 does not, and fails at the import with a message that reads like an auth problem.

Cost Optimization

• claude, agy, codex and copilot are subscription backends — none bills per call. api is the only metered path; reach for it only when you need agentic mode.
• Only extracts requested information (saves on output tokens)
• Avoids loading full files into main conversation context
• Use specific goals to minimize unnecessary processing

Troubleshooting

Issue	Solution
A CLI backend fails	Check that binary is installed and signed in: claude-code, agy, codex, copilot
codex blocks on stdin	The prompt must follow --; -i/--image is variadic and otherwise swallows it
Backend hangs or recurses	Confirm look_at.sh is the entry point — it sets LOOK_AT_NESTED=1; calling claude-code -p by hand does not, and image-read-guard.ts will then deny the child's Read
API key not set (api backend)	Set GOOGLE_API_KEY or GEMINI_API_KEY
File not found	Use absolute paths, verify file exists
Large file timeout	Break into smaller files or use lower-quality images
Rate limit errors	Add retry logic or use batch processing
Empty response	Check that goal is clear and specific

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