ClawPhone WeChat Control

处理微信会话列表、进入聊天、发送消息、处理微信内弹窗与聊天页失败排查。适用于用户要求查看微信消息、回复联系人、转发、处理聊天输入框或发送失败时。执行时必须先确认当前在微信的哪个页面，再按聊天场景一步一验。

Security Scan

The instructions clearly implement WeChat UI automation (enter chats, type/paste, click send) which matches the description, but the skill relies on powerful device-control primitives (screenshots, clipboard access, app-launch checks) without declaring those capabilities or describing safeguards — this raises privacy and data-exfiltration risk.

Purpose & Capability

The name/description and SKILL.md are coherent: the skill's instructions directly implement WeChat UI automation (launch_app('com.tencent.mm'), click_by_text, type_text, set_clipboard, screenshots, long-press/paste). However the metadata declares no required binaries, env vars, or permissions even though the runtime assumes device-control primitives (app launch, screenshot, clipboard, click_by_text, long-press). The lack of an explicit declaration of these capabilities/permissions is a mismatch worth noting.

Instruction Scope

Instructions repeatedly mandate taking screenshots and using the clipboard as verification/backup. Those actions can capture sensitive message contents and other on-screen data. SKILL.md does not limit what is captured, does not instruct how screenshots/clipboards are stored/retained/transmitted, and gives no user-consent or safety checks. While the steps stay within the stated task (WeChat automation), they create a high-risk surface for private data exposure.

Install Mechanism

This is an instruction-only skill with no install spec and no code files. That is the lowest install risk and consistent with the SKILL.md being runtime instructions only.

Credentials

The skill declares no environment variables or credentials, which is appropriate for a local UI automation instruction set. However it implicitly requires platform/device permissions (screenshot capture, clipboard write/read, ability to launch and interact with apps) that are not declared in metadata; the omission reduces transparency about what access will be needed.

Persistence & Privilege

always is false (good). The skill can be invoked autonomously (default) — combined with its need for sensitive device-level access this increases risk if the agent runs without explicit user confirmation. There is no instruction to require explicit, per-action user consent or to avoid uploading captured screenshots/clipboard contents.

What to consider before installing

This skill appears to do what it says: automate WeChat UI actions. However it instructs the agent to take screenshots and manipulate the clipboard (which can capture private messages) and to verify foreground app state — sensitive actions that could expose user data. Before installing: (1) confirm the agent runtime will require and enforce explicit device permissions (screenshot, clipboard, app control) and will only grant them with user consent; (2) require the skill to declare the exact capabilities it needs; (3) insist on safeguards: do not upload screenshots/clipboard content off-device, keep transient verification data ephemeral, log actions for audit, and require per-invocation user confirmation for reading or sending messages; (4) test in a non-sensitive account first. If those controls are not possible or you don't trust the runtime, do not enable this skill.

使用时机

• 用户要求读取或回复微信消息。
• 需要进入某个联系人聊天页。
• 聊天发送、转发、粘贴、长按菜单等操作失败。

页面判断

先区分当前处于哪一类页面：

• 桌面/非微信
• 微信会话列表
• 某个聊天页
• 联系人资料页
• 微信弹窗或菜单

不同页面的操作完全不同。不要在未确认页面时直接输入或发送。

回复消息标准流程

1. 确认已进入目标聊天页。
2. 截图确认底部真实输入框位置。
3. 只执行“点输入框”这一步，并再次确认键盘或光标已出现。
4. 先尝试 type_text(...)。
5. 若 type_text(...) 失败或文本未进入输入框，立即切换到“剪贴板 + 长按输入框 + 点粘贴”兜底路径。
6. 再次确认文本已经出现在输入框内。
7. 截图确认绿色发送按钮真实可见。
8. 点击当前截图里的发送按钮。
9. 再次截图，确认新的右侧消息气泡已经出现。

进入微信

1. 可先尝试 launch_app("com.tencent.mm")。
2. 立即检查前台应用是否真的变成 com.tencent.mm。
3. 若仍停留在桌面或其他 App，不要重复盲开；先读取当前桌面状态。
4. 桌面能识别到 微信 图标时，优先 click_by_text("微信") 进入。
5. 进入后再次确认当前页面是会话列表、聊天页还是弹窗页。

关键约束

• 不要在未聚焦输入框时直接 type_text(...)。
• 不要把某次截图里的输入框或发送按钮坐标复用到下一次。
• 不要把 press_enter 当成默认发送方式；只有可见发送按钮不可用时才可兜底尝试。
• 如果误入朋友资料页，先返回聊天页，再重新开始发送流程。

会话定位

• 会话列表优先用联系人文字进入。
• 若点击联系人失败，先截图确认是否有遮挡层、导入提示、搜索页或资料页。
• 若 click_by_text 无法命中会话项，但截图已能明确识别目标会话所在行，可基于当前截图临时点击该行；点完立即复核是否进入聊天页。
• 同名联系人或列表状态不明时，先截图确认，不要盲点。

输入兜底

• 输入框已聚焦，不代表 type_text(...) 一定能命中微信输入框。
• 若 type_text(...) 返回失败，优先执行：
  1. set_clipboard(...)
  2. 长按当前截图里的输入框空白区域
  3. 截图确认微信自定义菜单出现
  4. 基于当前截图临时点击 粘贴
• 粘贴后必须再次确认文本已进入输入框，再进行发送。

菜单与弹窗

• 微信自定义菜单通常不能靠 click_by_text 命中。
• 长按后先截图，再基于当前弹窗临时点击。
• 一步菜单一张图，不要连续盲点。

失败复盘模板

失败时按这几个问题复盘：

1. 当时在哪个页面。
2. 失败发生在“进聊天 / 聚焦输入框 / 输入文字 / 点击发送 / 发送验证”的哪一步。
3. 当时依据的是当前截图，还是误用了旧坐标/旧假设。
4. 是否做了操作后的二次确认。
5. 是否及时切换到了 launch_app、click_by_text、type_text 的兜底分支，而不是在同一路径上重复试错。