菜单
Use when you cannot inspect media directly and the user provides images, screenshots, image-heavy PDFs, or video frames. Send each image blob to the configured multimodal alternate, read its response, then continue using that response as context.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ahkohd/dotfiles --skill multimodal命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Getting Liquid Glass to render its specular rim (and its shadow) on macOS 26 — active appearance, tint rules, nesting rules, clipping rules, and shadow breathing room. Use when glass looks flat/rimless, when a glass panel or tooltip lacks the edge highlight the Dock's tooltips have, when a glass shadow clips at a window or view edge (visible on light backdrops), or when building any NSPanel/overlay/tooltip/toast/card that uses glassEffect or NSGlassEffectView.
Guide for the C -fbounds-safety language extension. Covers the language model, pointer annotations, adopting bounds-safety in existing C code, compiler build settings and modes, and runtime debugging of bounds violations.
Apple's complete prompt stack for Xcode 27's coding agent, as shipped. Consult when designing or reviewing agent prompts, tool descriptions, planner/executor splits, edit-application strategies, or context-window assembly — this is a production system's actual wording, not advice about prompting. Covers: planner and executor system prompts (including a GPT-5-specific variant and a no-classify variant), reasoning vs non-reasoning variants, tool-assisted variants, an A/B test pair, three edit integrators (standard, new-code, fast-apply), retrieval query expansion, and the fourteen fragments used to frame current file, selection, snippets, issues and search results.
正在显示 SKILL.md
| name | multimodal |
| description | Use when you cannot inspect media directly and the user provides images, screenshots, image-heavy PDFs, or video frames. Send each image blob to the configured multimodal alternate, read its response, then continue using that response as context. |
image -> alternate -> context for you
IMG=/path/to/image.png
MIME=$(file -b --mime-type "$IMG")
B64=$(base64 < "$IMG" | tr -d '\n')
curl -fsS https://llm.victor.computer/v1/chat/completions \
-H 'Content-Type: application/json' \
-d @- <<JSON | jq -r '.choices[0].message.content'
{
"model": "alternate",
"reasoning_effort": "none",
"chat_template_kwargs": {"enable_thinking": false, "preserve_thinking": false},
"messages": [{
"role": "user",
"content": [
{"type": "image_url", "image_url": {"url": "data:$MIME;base64,$B64"}},
{"type": "text", "text": "Describe this image concisely and factually."}
]
}]
}
JSON
Read the response, then continue the user's task.
The alternate is vision-only. Do not send it audio. Use configured local speech-to-text. If none is available, say so.
Say the local alternate is unavailable. Ask whether to wait, use a cloud model, or continue from the user's description.
Do not silently send private media to cloud providers.