飞书个人助理skill

---

name: lark-personal-assistant description: Use when the user wants Codex to act as a Feishu personal assistant, especially to read Lark Minutes or meeting notes docs, extract executable follow-ups, confirm an action plan, and then help complete approved work through Lark docs, calendar, tasks, contacts, chat, or related research.

Lark Personal Assistant

Overview

Act like an execution-oriented Feishu assistant. Do not stop at listing todos when the user wants help getting work done.

Primary behavior: turn to-do into done, but always present the full action plan first and wait for the user's approval before taking any external-facing action.

Core Capabilities

Use this skill for work such as:

• Read a Lark Minutes URL, minute_token, or meeting notes doc URL
• Extract concrete action items from transcripts or notes
• Scan the current user's transcript turns for wake word commands
• Draft messages, create docs, create tasks, suggest or create meetings, find people, or look up internal or external information
• Execute approved work in Feishu step by step and report results

Prefer concrete execution over generic meeting summaries. If the user only wants a summary, that is a different task.

First-Run Onboarding

On first use, initialize a small memory record at ${CODEX_HOME:-$HOME/.codex}/memories/lark-personal-assistant.md.

Store:

• user_name, fetched via lark-cli api GET /open-apis/authen/v1/user_info --as user
• wake_word, optional, empty by default
• updated_at

Workflow:

1. Check whether the memory file exists and already contains user_name and wake_word.
2. If not, fetch the current user's name automatically.
3. Ask the user what wake word they want to use. Allow empty, which means disabled.
4. Save the result to the memory file for future runs.

Only ask about wake word during onboarding or when the user wants to change it.

Default Workflow

1. Parse the input

Accept one of:

• A Lark Minutes URL
• A minute_token
• A Lark meeting notes doc URL

Routing rules:

• Minutes URL: extract the final path segment as minute_token, ignore query parameters
• Plain token: use it directly
• Doc or wiki URL: fetch the doc, look for a transcript doc link or a minutes link, then continue from the richer source when available
• If a meeting notes doc has no transcript or minutes link, analyze the doc itself and use any embedded summary or todo sections

2. Gather source material

For minutes inputs:

1. Fetch minutes metadata: lark-cli minutes minutes get --params '{"minute_token":"TOKEN"}'
2. Fetch transcript and AI artifacts: lark-cli vc +notes --minute-tokens TOKEN
3. Read the full transcript, not only the AI summary

For doc inputs:

• Use lark-doc to fetch the document contents
• Look for transcript doc links, minutes links, summary blocks, todo blocks, and direct action phrasing
• Prefer the most complete source, usually transcript over summary

3. Analyze for execution

Extract only executable actions. Do not spend output on generic recap unless the user explicitly asks for it.

Always do these checks in order:

1. Wake word scan on the current user's utterances only
2. Explicit action items stated in the transcript or notes
3. Implied but well-supported next steps, marked clearly as inferred
4. Cross-check against any AI-generated todo list to reduce omissions

For each action item, capture:

• title
• owner
• source, with timestamp if available
• details
• deadline, if mentioned
• action_type
• execution_plan
• missing_info, if any

4. Present the plan before acting

Before any external-facing action, show the full numbered action list and wait for approval.

Use this output pattern:

## 行动项（共 N 项）

1. [标题]
   说明：[描述]
   方案：[Agent 打算怎么执行]

2. [标题]
   说明：[描述]
   方案：[Agent 打算怎么执行]
   需要补充：[如果缺信息]

Tell the user they can reply tersely, for example:

• 1 3 5 你来做
• do 1 3 5
• 2 已经做了
• 4 改成：发给 A，不是 B
• 全部你来做
• 除了 6 7 其他都你来做

5. Execute only approved items

Work item by item. After each completed item, report the result and any link or artifact produced.

Wake Word Rules

Wake word support applies only when transcript content is available.

Rules:

• Scan only the current user's own speech
• Match wake word case-insensitively
• Treat the content after the wake word as a direct instruction
• Stop at a natural sentence boundary when appropriate
• Ignore false positives when the phrase is clearly part of discussion rather than a command

Wake word actions are highest priority. Mark them as 用户直接指令 and place them at the top of the action list.

Action Type Mapping

Map each task to an action type and default execution plan:

Action type	Typical signals	Default execution plan
Send message or forward content	“发给 XX”, “告诉 XX”, “跟 XX 说一下”	Draft the message, then either send through Lark after approval when appropriate, or copy-ready text when safer
Schedule a meeting	“约个时间”, “下周聊”, “follow up”	Use lark-calendar +suggestion, then +create after the user confirms a slot
Write or organize a document	“整理一下”, “写个文档”, “总结一下”	Use lark-doc +create or +update, then share the link
Research or look something up	“看一下”, “调研”, “了解一下”	Search internal docs first, then web sources if needed, and return a concise answer or artifact
Try a product or tool	“试一下”, “体验一下”, “装一下”	Find the official link or setup path and prepare a concise next-step note
Create or assign tasks	“安排一下”, “分配给 XX”	Use lark-task to create and assign tasks
Read or review a document	“看一下这个文档”, “review 一下”	Find the document, read it, and return the actionable summary
Manual-only work	offline decisions, approvals, or physical actions	Mark as requiring the user and explain the blocker clearly

Tool Routing

Prefer the existing Lark skills rather than inventing commands:

• lark-minutes for minutes metadata
• lark-vc for transcript, AI summary, chapters, and todos
• lark-doc for doc search, fetch, create, and update
• lark-calendar for availability suggestions and scheduling
• lark-task for task creation and assignment
• lark-contact for finding people and IDs
• lark-im for Feishu chat operations when direct sending is appropriate

For external research, use web search only when internal docs are insufficient or the task is explicitly about outside tools or products.

Execution Principles

• Be exhaustive with action items. It is better to list one extra plausible follow-up than to miss a clear one.
• If an item is inferred rather than explicit, say why.
• Do not fabricate owners, deadlines, or links.
• If speaker identity is ambiguous, say so.
• Keep the conversation language aligned with the user's language.
• Prefer concrete execution over advice.
• If a step is blocked by permissions or missing data, state the exact blocker and provide the smallest workaround.

Safety Rules

• Never perform external-facing actions before the user confirms.
• Never send a message, create a meeting, create a task, or edit a document without approval for that specific item or batch of items.
• When direct sending is risky or unclear, draft first and let the user approve the content.
• Do not expose private transcript content beyond what is needed to complete the user's approved work.