Describe images (screenshots, error screens, UI mockups, charts) in detail when the current model cannot see images. Use when the task involves screenshots, images, UI mockups, charts, or error screenshots
Installs into .claude/skills of the current project.
Are you the author of Deepseek Vision?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/limccn-deepseek-vision)
---
# deepseek-vl-support:skill
name: deepseek-vision
description: Describe images (screenshots, error screens, UI mockups, charts) in detail when the current model cannot see images. Use when the task involves screenshots, images, UI mockups, charts, or error screenshots
allowed-tools: Bash Read
---
# deepseek-vision — Describe images
The current model (e.g. DeepSeek) is text-only and cannot see images. When the
task requires understanding an image file, use this skill to obtain a precise
text description and reason from it.
## When to use
- error screenshots
- UI mockups and interface screenshots
- charts and diagrams
- any image the model needs to understand
## How to use
Run in the Bash tool:
```bash
npx @limccn/deepseek-vl-support describe "<image file path>" "<optional question>"
```
or, if the `deepseek-vl-support` bin is on PATH:
```bash
deepseek-vl-support describe "<image file path>"
```
The command is identical in bash, zsh (macOS default), and PowerShell
(Windows): quote paths containing spaces with `"`. No shell-specific
syntax or call operator is needed — `npx` resolves via its `.cmd` shim
under PowerShell.
Read the returned description and use it for reasoning. Large images (over
`maxBytes`, default 10 MB) are rejected with a clear error — compress or crop
first (e.g. under 5 MB, ~2000px long edge).
## Customizing the vision prompt
Default prompt: `references/vision-prompt.md`. Overrides (highest wins):
- project `.deepseek-vl/vision-prompt.md`
- global `~/.deepseek-vl/vision-prompt.md`
## Troubleshooting
Run `npx @limccn/deepseek-vl-support doctor` to check the vision endpoint and model.
If images are not described automatically, check `VISION_DISABLE` and the hook
registration in `.claude/settings.json`.
