Lark Doc スキル

A sourcing note before the quotes: this skill has no public repository. The canonical file is served at open.feishu.cn/.well-known/skills/lark-doc/SKILL.md, which we fetched on August 20, 2026 and verified matches our catalog copy exactly. There is no commit history to pin, so quotes cite the served file and that verification date. Translations from the Chinese are ours, and where the file uses inline link markup we quote the rendered text (the link's label) rather than the raw markdown.

lark-doc at a glance

1. The skill is written in Chinese and leads a 27-skill suite whose siblings install together.
2. It declares hard dependencies in frontmatter: a lark-cli binary and a sibling skill, before a single instruction runs.
3. It is a context-budget router, whose loudest rule forbids reading everything up front.
4. "A simple task is not a reason to skip": the one discipline line that separates it from most workflow skills.
5. Auth loads only on failure, a lazy-error pattern most skills get backwards.

1. The skill is written in Chinese and leads a 27-skill suite

The file is Chinese top to bottom, a stray loanword aside (its H1 is a plain # docs, and "CRITICAL" survives in English). The description opens:

飞书云文档（Docx / Wiki）内容操作：读取、创建、编辑文档，插入或下载图片附件，以及操作思维笔记。

Source: SKILL.md description, served at open.feishu.cn, verified 2026-08-20

Our translation: "Feishu cloud document (Docx / Wiki) content operations: read, create, and edit documents, insert or download image attachments, and work with mind notes." The language costs the agent little and the human auditor a lot: you cannot skim what you cannot read. The suite fact matters just as much: open.feishu.cn publishes 27 skills, and 22 of them sit within about 2,000 installs of lark-doc's 877k (Skillselion catalog, August 20, 2026), the signature of tools installed together as a bundle rather than picked one by one. One routing detail worth knowing: the description tells the agent to route by URL path and token, not by domain, and explicitly includes doubao.com document links, ByteDance's consumer AI product.

2. It declares hard dependencies in frontmatter

Before any instruction, the metadata names what must already exist:

metadata: requires: bins: ["lark-cli"] skills: ["lark-shared"]

Source: SKILL.md frontmatter, served at open.feishu.cn, verified 2026-08-20

A binary and a sibling skill, declared as machine-readable requirements rather than prose. Compare the shims we covered in grill-me, where the dependency on the grilling skill is implicit in a sentence: here the dependency graph is explicit metadata a tool can check before install. lark-doc is one node in a suite; the file hands file management and comments to lark-drive and whiteboard writes to lark-whiteboard, each its own skill.

3. It is a context-budget router

The first body instruction is bolded, capitalized, and about token spend:

CRITICAL：先判断场景，再读取该场景的参考文件；不要在任务开始时一次性读取全部参考文件。每个文件只在首次进入对应阶段时读取一次。

Source: SKILL.md, "场景与 Shortcut 路由", served at open.feishu.cn, verified 2026-08-20

Our translation: "CRITICAL: judge the scenario first, then read that scenario's reference file; do not read all the reference files at once when the task starts. Read each file only once, on first entering its stage." The SKILL.md itself is a 50-line routing table pointing at a dozen reference files (fetch, create, update, history, media, whiteboards, mind notes). The design goal is the same progressive disclosure that vercel-react-best-practices gets from its rules index: the always-loaded file stays small, and detail is paid for only when the task needs it.

4. "A simple task is not a reason to skip"

The routing entry for writing a document from scratch carries the file's best line:

从零创作 — 创建工作流：先完整执行创建工作流，简单任务不是跳过的理由；

Source: SKILL.md, "文档内容", served at open.feishu.cn, verified 2026-08-20

Our translation: "Creating from zero: execute the creation workflow in full first; a simple task is not a reason to skip it." Whoever wrote that has watched an agent decide a task was too small for the process and ship something malformed. The escape hatch is explicit and narrow: only importing user-provided content verbatim, or creating an empty document, skips the workflow.

5. Auth loads only on failure

Most skills front-load authentication setup. This one forbids it:

执行 Shortcut 时，不预读 lark-shared 或预跑 auth status --verify；仅遇到未认证、token / 身份或 scope 错误时读取该 Skill，修复后重试。

Source: SKILL.md, "认证与 Scope", served at open.feishu.cn, verified 2026-08-20

Our translation: "When executing a shortcut, do not pre-read lark-shared or pre-run auth status --verify; read that skill only when you hit an unauthenticated, token/identity, or scope error, fix it, and retry." The optimistic path costs nothing when auth is already valid, which it usually is. It is the same lazy philosophy as section 3 applied to error handling: pay for the fix only when the failure is real.

Tools it drives: the lark-cli binary (docs and mindnotes subcommands), against Lark/Feishu's document APIs.

Behavior change, before and after:

The fine print worth knowing:

• The skill is scoped tightly on purpose: spreadsheets and Base tables are out, file-level operations and comments belong to lark-drive. If your team lives in Lark, install the suite, not just this node.

Should you install lark-doc?

Install it if your team's documents live in Lark or Feishu; nothing else in the catalog gives an agent this surface, and the file's context discipline is genuinely well engineered. Skip it entirely if you do not use Lark: every capability here assumes ByteDance's ecosystem, from the CLI to the URL formats it routes on.

npx skills add https://open.feishu.cn --skill lark-doc

Previous: how vercel-react-best-practices actually works. All dives: the series index.