Communication Assistant

Use this skill as the routing layer for personal communication tasks across chat and email. It does not replace platform-specific skills; it coordinates them, keeps a shared outbox, and enforces a conservative "prefill, do not send" default.

Defaults

• Keep normal summaries in Chinese.
• Never look for, recover, ask to store, or record account passwords.
• Use QR login, existing browser/client sessions, official OAuth/API tokens, or platform-local session files only.
• Store any sessions, tokens, configs, and message outbox data in the git-ignored runtime folder described below.
• Do not expose private chats, contacts, tokens, cookies, QR codes, or message bodies in public wiki pages, commits, or partner prompts.
• Prefer prefilled over sent. Real sending requires explicit user approval for the exact platform, target, and outbox item.
• If a platform is blocked or unsafe, mark that adapter as pending instead of pretending it works.

Runtime Paths

Runtime data lives under $env:COMMUNICATION_ASSISTANT_HOME if it is set. Otherwise it goes to .wiki-tmp/communication-assistant/ inside the skill folder (or inside the enclosing notes repo, if the skill sits in a folder that has both AGENTS.md and .wiki-schema.md). Either way it has four git-ignored subfolders: config, sessions, outbox and logs.

The helper script prints the folders in use and manages outbox items. Run it from the skill folder:

node scripts/outbox.js paths
node scripts/outbox.js create --platform wechat --target "Alice" --message "我晚点回复你。" --context "reply to unread message"
node scripts/outbox.js list
node scripts/outbox.js mark --id "<outbox-id>" --status prefilled
node scripts/outbox.js cancel --id "<outbox-id>"

Platform Routing

Use the safest available adapter for each task:

Read references only when needed:

• references/adapters.md for adapter selection and setup notes.
• references/safety.md for approval, secret, and live-account boundaries.
• references/outbox.md for the shared draft/prefill lifecycle.

Workflow

1. Identify platform and account/session.
2. Read or search through the least intrusive safe adapter.
3. Summarize in Chinese:
   • who sent it
   • what they want
   • urgency or deadline
   • risk / hidden obligation
   • suggested response
4. Draft the response and create an outbox item.
5. If the user asks to prefill, place text in the chat input without sending when the adapter supports it.
6. If the user explicitly approves sending, verify target and text one more time, then use the platform adapter's send path.

Useful User Prompts

帮我看微信未读消息，中文讲重点。
帮我回 WhatsApp 这条，先别发。
查一下 QQ 有没有课程/签证/账单相关消息。
把刚才那条回复填到微信聊天框，不要发送。
我确认，发送 outbox 里这条。

Safety Checks Before Live Actions

• Confirm the platform and target.
• Confirm whether the action is read-only, outbox-only, prefill-only, or send.
• For send, require explicit approval in the current conversation.
• Do not use delegated partners for raw private chat content unless the user explicitly approves sharing that content.
• When browser or desktop automation is needed, warn that a window may open or focus may change.