| name | image-vision |
| description | Look at what's actually in an image — describe the scene, subjects, colors, mood, text, composition. Use when the user attaches a photo and asks what's in it, to critique it, or anything that requires seeing the picture content (not just its file facts). |
| allowed-tools | Bash(python3:*) |
Image vision
When the user attaches an image, the message includes a marker like:
[image: photo.jpg — saved to /workspace/agent/attachments/playground_xxx_0.jpg]
That path is the image file on disk. To actually SEE the picture, run the
bundled script with the path and (optionally) a question:
python3 /app/skills/image-vision/scripts/describe.py "/workspace/agent/attachments/playground_xxx_0.jpg" "Is this a strong ad photo? What would you change?"
The script sends the image to a vision model and prints back a description /
answer. With no question it just describes the image.
Resolution / cost knob — add --detail:
python3 /app/skills/image-vision/scripts/describe.py "<path>" "Read the text on the sign." --detail high
--detail low — model sees ~512px. Cheapest and fastest; fine for overall
scene, colors, mood.
--detail high — image is tiled at full resolution. More tokens (costs more)
but needed for small text, fine detail, or close inspection.
--detail auto (default) — the model decides.
Use low for quick "what's this?" and high when the user asks about text in
the image or fine details. (--model NAME overrides the vision model.)
Use it for "what's in this photo?", "critique this ad image", "what colors
dominate?", "is there text in it?", "does the composition work?" — anything
that needs looking at the actual pixels.
For plain file facts (dimensions, aspect ratio, file size) you don't need
vision — use the Image info skill, which is faster and needs no network.