image-analyst

name

image-analyst

description

Use this skill when the user uploads an image, screenshot, chart, diagram, whiteboard photo, design mockup, document photo, error screenshot, or asks Prometheus to visually inspect a local image file. Triggers on phrases like analyze this image, what is in this screenshot, read this chart, extract text from this image, inspect this mockup, diagnose this error screenshot, OCR this photo, compare these images, and what do you see here. Use it to describe, extract, interpret, and recommend next actions from visual evidence.

Image Analyst

Use this skill when images have been uploaded to the chat or when the user wants to analyze a visual file. Works best with vision-capable providers (OpenAI, OpenAI Codex).

---

1. When to Use

User input	Analysis type
Screenshot of app/website	UI element inventory, layout critique, UX issues
Chart or graph	Data extraction, trend reading, axis/label parsing
Diagram (flowchart, architecture, ER diagram)	Element identification, relationship mapping
Handwritten notes or whiteboard photo	OCR text extraction, idea summary
Design mockup	Feedback, component list, style guide extraction
Photo of a document	Text extraction, key information summary
Error screenshot	Error message extraction, diagnosis
Any image + a question	Answer the question using visual context

---

2. Analysis Framework

For every image, deliver in this order:

A. Describe (What am I looking at?)

• Type of image (screenshot, chart, diagram, photo, etc.)
• Overall subject/context in one sentence
• Dimensions or rough aspect ratio (if relevant)

B. Extract (What does it contain?)

Pull out all observable data:

• Text — every visible string, label, heading, value, placeholder, error message
• UI elements — buttons, inputs, dropdowns, nav items, modals (for screenshots)
• Data values — chart Y-axis values, percentages, table rows, key metrics (for charts/tables)
• Structure — sections, panels, boxes, connections, hierarchy (for diagrams)
• Colors/style — primary palette, typography cues (for design mockups)

C. Interpret (What does it mean?)

• What is the main takeaway or insight?
• For charts: what trend or comparison is being shown?
• For UI: what is the user expected to do here?
• For diagrams: what is the system/process being described?
• For error screenshots: what went wrong and likely why?

D. Act (What should happen next?)

Provide specific, actionable output based on the user's request:

• Answer their specific question if they asked one
• List issues found (bugs, UX problems, data anomalies)
• Suggest improvements or next steps
• Extract structured data (tables, JSON, markdown) if that's the goal

---

3. Output Templates

For UI/Screenshot Analysis

## Image Analysis: [Page/Component Name]

**Type:** App screenshot — [web/desktop/mobile]
**Summary:** [One sentence about what this shows]

### Elements Found
- Navigation: [items]
- Key content: [headings, body copy]
- Interactive: [buttons, inputs, CTAs]
- Status indicators: [errors, badges, notifications]

### Key Observations
1. [Most important thing you see]
2. [Second observation]
3. [Third observation]

### Issues / UX Notes
- [Issue or improvement opportunity]

### Recommended Action
[What to do based on this screenshot]

For Charts/Data

## Chart Analysis

**Chart type:** [Line, bar, pie, scatter, etc.]
**X-axis:** [label and range]
**Y-axis:** [label and range]
**Data series:** [series names]

### Key Values
| [X] | [Y] |
|---|---|
| [val] | [val] |

### Trend/Insight
[Main takeaway in 1-2 sentences]

For Diagrams

## Diagram Analysis

**Type:** [Flowchart / ER / Architecture / Mind Map / etc.]
**Subject:** [What system or process]

### Components
- [Component A] — [role]
- [Component B] — [role]

### Relationships
- [A] → [B]: [relationship description]

### Summary
[One paragraph describing what this diagram communicates]

---

4. OCR Mode (Reading Text from Images)

When the goal is just to extract text:

1. Read every visible string, left to right, top to bottom
2. Preserve formatting cues (headings look bigger → use ## in output)
3. Mark uncertain text with [?]
4. Preserve numbers exactly (don't round or infer)
5. Wrap extracted text in a code block for easy copying

---

5. Vision Limitations

Be honest about:

• Low resolution: Mark blurry/pixelated text as [unclear]
• Cut-off content: Note if image appears cropped
• Color-blind safe: Describe charts by shape/position, not just color
• Dynamic content: Screenshots are static — note if time-sensitive data shown

---

6. Multi-Image Workflow

If multiple images are uploaded in one message:

1. Number them: "Image 1:", "Image 2:", etc.
2. Analyze each separately under its own heading
3. Then synthesize: "Comparing all images..." section at the end
4. Note relationships or contradictions between images