Image Reading

Use this skill when the job is to understand what an image says, not just to describe that an image exists.

Good fits:

• screenshots
• app or website UI captures
• posters, ads, menus, flyers, slides
• receipts, invoices, labels, forms
• charts, dashboards, scorecards, tables
• social posts or meme screenshots where both text and layout matter

Core Rule

Do two jobs, in this order:

1. Read what is visibly there.
2. Explain what it means.

Never blur the line between those two.

Workflow

1. Inspect the image directly

If the user provided a local image path, load it with the image-viewing tool first.

Use direct visual inspection to capture:

• layout
• sections
• hierarchy
• colors or emphasis that change meaning
• icons, charts, buttons, UI states, or branding
• anything OCR alone will miss

2. Use OCR when exact text matters

If an OCR tool is available and the image is text-heavy, run OCR after visual inspection.

Use OCR especially for:

• dense screenshots
• receipts or forms
• small text
• tables
• long paragraphs
• timestamps, numbers, totals, labels, or IDs

Do not trust OCR blindly. Cross-check obvious mistakes against the image.

3. Reconstruct the content cleanly

Extract the visible content in a human-useful order:

• top to bottom
• left to right
• grouped by section if the image has clear panels or blocks

If the image is structured, preserve that structure instead of flattening everything into one paragraph.

4. Separate fact from inference

Always keep these apart:

• what the image explicitly says
• what you infer from the image
• what is uncertain or unreadable

If a word, number, or label is not fully legible, say so.

5. Answer at the right level

If the user asks:

• "what does this say?" -> extract the content first
• "what does this mean?" -> explain the meaning after extraction
• "summarize this" -> summarize the content while preserving important numbers or claims
• "is there anything important here?" -> surface the high-signal points, warnings, deadlines, totals, or risks

Response Contract

Use the lightest shape that fits the task. When the image is important or dense, prefer:

1. Visible text
2. What it shows
3. What it likely means
4. Unclear or uncertain

For charts, dashboards, or tables, prefer:

1. What is on the screen
2. Key numbers or labels
3. Trend or takeaway
4. Unclear or missing context

For receipts, bills, and forms, prefer:

1. Document type
2. Important fields
3. Totals / dates / names / IDs
4. Anything unusual or unclear

Rules for Accuracy

• Quote short exact text when useful, but do not pretend uncertain OCR is certain.
• Keep numbers exact when they are legible.
• If you infer something from layout, branding, or context, label it as inference.
• Do not invent hidden text.
• Do not over-interpret charts without naming the visible evidence first.

Special Cases

UI screenshots

Capture:

• page or app state
• active panels or tabs
• errors, warnings, banners, toggles, or missing states
• the main action the user can or cannot take

Social screenshots

Capture:

• the post text
• account or source if visible
• engagement counts or timestamps if relevant
• whether replies, images, or quoted content change the meaning

Charts and dashboards

Capture:

• title
• axes
• labels
• legends
• headline values
• obvious trends

If context is missing, say the trend is visible but the cause is not proven.

When To Be Brief

If the image is simple and the user asks a direct question, answer directly instead of forcing the full structure.

When To Slow Down

Slow down when the image contains:

• money
• deadlines
• legal or medical wording
• fine print
• multiple panels
• hard-to-read small text
• data that could be misread