ta
Use the Ta macOS app and ta CLI to capture displays, foreground apps, windows, or known regions without taking focus, then annotate, OCR, understand, translate, copy, or save images. Use when an Agent needs visual input or deterministic screenshot markup; do not use it as a general mouse-and-keyboard automation tool.
中文简介正在准备中。当前显示原文说明。
查看安装方法包含的文件(9)
- SKILL.md3.3 KB
- agents/openai.yaml196 B
- references/capture-targeting.md1.7 KB
- references/commands.md3.4 KB
- references/editing-recipes.md4.2 KB
- references/error-codes.md2.0 KB
- references/image-understanding.md1.5 KB
- references/privacy-and-safety.md1.7 KB
- scripts/check-ta.sh406 B
SKILL.md(原文)
安装前可以查看交给 AI 代理的指令内容。
Ta Screen Intelligence
Use Ta as the visual input layer. Ta owns macOS Screen Recording permission, model profiles, and secrets; the CLI only sends requests to its local bridge.
Start safely
- Run
scripts/check-ta.shonce before the first Ta action in a task. - Read the JSON response. Continue only when top-level
okistrueanddata.bridgeisready. - Use
--jsonfor every command an Agent must inspect. Treat a request as successful only when top-levelokistrue; image-producing commands must also return a non-emptyartifactsarray. - Do not request, read, print, or pass API keys. Model configuration stays in Ta.
If the check fails, follow references/error-codes.md. Do not repeatedly launch Ta or retry non-retryable errors.
Choose the workflow
- Capture the current page or app:
ta capture frontmost --json. Ta freezes the foreground target and should not activate itself, move the pointer, or send input events. - Capture a particular visible window: list candidates first, then capture the returned ID. Read references/capture-targeting.md.
- Extract text: capture, then run local OCR. Read references/image-understanding.md.
- Understand semantics, code, tables, or formulas: use
ta analyzeonly when OCR is insufficient. This may use the configured cloud model. - Translate visible content: choose text translation for a textual result or image translation for a rendered translated image. Read references/image-understanding.md.
- Copy or save a result: do so only when the request calls for changing the clipboard or filesystem. Read references/commands.md.
- Annotate an image reproducibly: use
ta transform ... --recipe <JSON>and keep stable IDs for later erasing. The transform is local and returns a new PNG Artifact. Read references/editing-recipes.md. - Beautify, redact, pin, or long-capture: check
ta capabilities --jsonfirst. Do not invent methods absent from the returned list.
Privacy boundary
Ordinary screenshot requests are pre-authorized after the user grants Ta macOS permission. Still stop before capturing a password manager, banking/payment app, private-key or recovery-code screen, private browsing window, or another obviously sensitive target unless the user explicitly identifies that target for capture.
Use local OCR when text alone is enough. Do not silently switch a cloud-denied task to cloud processing or an unavailable silent long capture to an interactive flow. Read references/privacy-and-safety.md for cloud, clipboard, and artifact rules.
Report observable results
Report the target actually returned by Ta, whether cloud upload occurred, and the resulting text or artifact path. Temporary artifacts expire; save a durable copy when the user needs the image later. Never claim success from command dispatch alone.
评价
还没有评价。欢迎分享你的使用感受。
评论
无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。
还没有评论。来写第一条吧。