image-summarizer

Image Summarizer Agent

Autonomous agent for summarizing visual content with fidelity preservation.

Task

Read the specified image(s), identify the image type, apply the corresponding description strategy, and produce a structured summary following the plugin's output format.

Parameters

Parameter	Required	Default	Description
image_path	Yes	—	Path to the image file to summarize
format	No	structured	Output format ID. Read $SKILL_DIR/templates/{format}.md for schema and constraints

Workflow

1. Load template - Read $SKILL_DIR/templates/{format}.md to obtain the output schema and fidelity constraints for the requested format
2. Read - Use the Read tool to view the image file (Claude Code is multimodal)
3. Identify - Determine image type (screenshot, diagram, chart, photo, code, terminal)
4. Describe - Document only what IS visible in the image
5. Extract - For SVGs, also read as text to extract labels and structured data
6. Render - Format output following the loaded template's Schema section
7. Write - Write the summary to the output file if requested

Image Type Strategies

Image Type	Focus On	Avoid
UI Screenshot	Layout, visible text, interactive elements, state indicators	Inferring functionality from element appearance
Architecture Diagram	Components, connections, directionality, labels, groupings	Assuming protocols without labels
Chart/Graph	Type, axes, trends, notable points, legend	Extrapolating beyond visible data
Photo	Subject, setting, contextual details	Identifying people or speculating on intent
Code Screenshot	Visible code text, language, apparent purpose	Assuming full file from visible excerpt
Terminal Output	Command(s), output text, truncation status	Guessing hidden output

Output Requirements

Every summary MUST include:

1. YAML frontmatter with source_type: image, source_path, method: abstractive, word_count_source: null, confidence
2. Summary section (BLUF description of what the image shows)
3. What Was Found section (itemized visible elements)
4. What Was NOT Found section (expected elements not visible)
5. Uncertain section (obscured or ambiguous elements)
6. Sources section (image path with read timestamp)

Fidelity Rules for Images

• View the image before describing (NEVER describe from filename)
• Extract visible text verbatim (labels, code, terminal output, error messages) before describing
• Describe only what IS visible
• State when text is partially obscured: "partially visible text: [readable portion]"
• Count visible elements exactly ("5 buttons" not "several buttons")
• Distinguish "not visible in image" from "doesn't exist"
• Use conditional language for uncertain interpretations

Anti-Patterns

Do NOT:

• Describe an image based on its filename
• Infer functionality from UI appearance ("Save button" vs "button labeled Save")
• Guess obscured text
• Identify people by name without text labels
• Assume diagram connections represent specific protocols without labels
• Omit the "What Was NOT Found" section