ArmorandClaude Opus 4.7
8d281eefd6
- scripts/vision.py: call MiMo v2.5 via Anthropic-format API for image analysis - scripts/screenshot_mission_list.py: Playwright screenshot of /missions page - .claude/skills/vision.md: Claude Code skill for vision task delegation Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
46 lines
1.6 KiB
Markdown
46 lines
1.6 KiB
Markdown
---
|
|
name: vision
|
|
description: Use when you need to visually inspect a screenshot, UI render, or any image — the current model has no multimodal capability, so vision tasks MUST be delegated to this tool.
|
|
---
|
|
|
|
# Vision Tool
|
|
|
|
When you need to look at an image (screenshot, UI render, diagram, photo), call the MiMo v2.5 multimodal model via this script.
|
|
|
|
## Usage
|
|
|
|
```bash
|
|
python scripts/vision.py "<image_path>" "<question>"
|
|
```
|
|
|
|
## When to Call
|
|
|
|
You MUST use this for any of the following:
|
|
- Checking a UI screenshot for layout issues (overlapping elements, overflow, alignment)
|
|
- Verifying a page render looks correct after a change
|
|
- Reading text or data from a screenshot
|
|
- Analyzing any image the user provides
|
|
- Any task where you think "I wish I could see this"
|
|
|
|
## Examples
|
|
|
|
```bash
|
|
# Basic UI inspection
|
|
python scripts/vision.py "demo/test_screenshots/timeline_page.png" \
|
|
"Check this timeline page for any UI issues: overlapping elements, clipped text, misaligned segments. Report each issue with its location."
|
|
|
|
# Comparing two screenshots
|
|
python scripts/vision.py "demo/test_screenshots/docking_modal.png" \
|
|
"Is the docking modal properly centered? Are all buttons visible and correctly labeled? Is the datetime input functional-looking?"
|
|
|
|
# Reading wiki page render
|
|
python scripts/vision.py "data/KSP/Vulture Shuttle Full Stack at Launch Pad.png" \
|
|
"Describe what you see in this image in detail."
|
|
```
|
|
|
|
## Notes
|
|
|
|
- Images over ~5 MB may be slow; consider resizing large screenshots first
|
|
- The MiMo model runs remotely — no local GPU needed
|
|
- Response is text-only; the model describes what it sees
|