claude-vision-skill
Use when the user shares, pastes, or references an image (local path or URL) and you need to describe, analyze, or recognize its content, especially when the current model cannot read images directly. Run the bundled vision.js helper to convert the image into text.
How do I install this agent skill?
npx skills add https://github.com/asuojun/claude-vision-skill --skill claude-vision-skillIs this agent skill safe to install?
- Gen Agent Trust Hubfail
The skill enables vision capabilities by sending local images to external APIs. It requires bypassing system execution policies to handle clipboard data and instructs the agent to modify its own host application source code and configuration files, which can lead to unauthorized persistence and behavioral changes.
- Socketwarn
1 alert: gptAnomaly
- Snykwarn
Risk: MEDIUM · 1 issue
What does this agent skill do?
Vision Helper
The current model may not support native image input. This skill is a sidecar workflow: it sends the image to a separate vision model and returns text to the current model. It does not add native image input to the current model.
When this skill is installed as a Claude Code skill, use the bundled script
through ${CLAUDE_SKILL_DIR} so the command works from personal, project, or
plugin installations:
node "${CLAUDE_SKILL_DIR}/vision.js" "<absolute image path>" "<prompt>"
If the repository was installed using the project-root setup described in the README, use the copied script instead:
node vision.js "<absolute image path>" "<prompt>"
For an image URL:
node "${CLAUDE_SKILL_DIR}/vision.js" --url "<image url>" "<prompt>"
Use node vision.js --url instead when using the project-root setup.
When the user pastes an image into the chat but no file path or URL is visible, do not pass the raw image to the text-only model. Read the image from the clipboard first:
node "${CLAUDE_SKILL_DIR}/vision.js" --clipboard "<prompt>"
Use node vision.js --clipboard instead when using the project-root setup.
--clipboard reads the current image from the system clipboard (macOS uses a bundled Swift helper, Windows uses a bundled PowerShell script; the pasted image is usually still there). If it fails, ask the user to save the image to a file and provide the absolute path.
Fallback rules (automatic):
- If a local path is given but the file does not exist, vision.js automatically falls back to the clipboard.
- If no image path or URL is given at all, vision.js automatically tries the clipboard.
- Pass
--no-fallbackto disable this behavior and fail with an explicit error instead.
Rules:
- Use
${CLAUDE_SKILL_DIR}/vision.jsfor a bundled skill installation, or the project-rootvision.jsfor the manual setup. Never use a path tied to the maintainer's machine. - Use an absolute image path for local files, or
--urlfor remote images. - Prefer
--clipboardwhen the user pasted an image with no accessible path. - Pass the user's original question as the prompt instead of replacing it with a generic description request. Use Chinese for descriptions unless the user asks otherwise.
- Configuration lives in
.envnext tovision.js(DASHSCOPE_API_KEY,VISION_MODEL,DASHSCOPE_BASE_URL). Never print or commit the API key. - If the API call fails, report the error to the user and ask them to check the key, model, or base URL.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/asuojun/claude-vision-skill/claude-vision-skill">View claude-vision-skill on skillZs</a>