Lark Skills Roundup: Slides XML, Task GUIDs, Whiteboard CLI, OKR Locks, Minutes Shim

Feishu’s lark family still has five skills this site had not cracked open, so I read every file in the repo and scraped skills.sh on 2026-09-30. The pack counters sit at 737.4K–737.8K installs for lark-minutes, lark-slides, lark-task, and lark-whiteboard (all first seen Apr 14 2026), with lark-okr lagging at 718.1K (first seen Apr 17). Weekly installs locked step again: across all eight buckets the five track in an 81–229 spread, tightest on the oldest week — the pack-counter signature lark-im, lark-sheets, and lark-base showed earlier. The file sizes tell the real story — a 464-byte shim next to a 27.8KB XML protocol — so here is what each one actually forces an agent to do.

The 15-Line Shim: lark-minutes#

lark-minutes is the thinnest file in the family: 15 lines, 464 bytes of frontmatter and a routing note. Its description says use it only when the user or an upstream config explicitly names lark-minutes; everything else routes to lark-meeting. The body demands a full read of ../lark-meeting/SKILL.md and execution per that skill’s routing. Same compatibility pattern as lark-vc-agent’s 466-byte shim — the family keeps old names alive as pointers, not copies. If your config references lark-minutes and you see “Minutes” versus “Notes” confusion, the answer is in the meeting skill, not here: minute_token is its own entity, and local audio can generate Minutes without any meeting attached.

Task GUIDs and the Search-Vs-List Trap: lark-task#

lark-task (v2, 13.6KB) opens with a progressive-discovery rule: never construct a +verb from guesswork. Match the intent against the shortcut table; no match means lark-cli task --help, then the exact token, then schema before any call. An unknown_subcommand error means stop guessing and re-discover.

Two traps will bite automation. First, the guid used by every update/complete/assign call is the global unique identifier — the one in the guid= query of a task applink — not the displayed task number (t104121, suite_entity_num). Feed the display number to +update and you touch the wrong object. Second, +search does not judge relevance to the caller: “search my tasks” needs the current user’s open_id resolved and passed explicitly as --assignee, --creator, or --follower, or you get everyone’s matches. Scope words are not queries either — “tasks this year” is a list scope (+get-my-tasks / +get-related-tasks), not a full-text search, and time expressions must never become the query value. The skill even arbitrates the word “todo”: if the context is a meeting transcript or a minute_token, minutes +todo belongs to lark-meeting and task +create is forbidden. One more rule: repeat_rule and reminder only exist when due is set, and start must be ≤ due.

Slides Runs on an XML Protocol: lark-slides#

lark-slides (v1, 27.8KB, 310 lines) is the heavyweight — the skill tells you to read it twice. Slides live on a 960×540 canvas and editing means writing slide XML: references/xml/slides_xml_schema_definition.xml is the only protocol authority. New decks and large rewrites require a .lark-slides/plan/<deck>/slide_plan.json planning artifact first, and every <slide> XML must pass scripts/xml_lint.py with summary.error_count === 0 before any API call.

The edit verbs map to damage levels: +replace-slide does block-level replace/insert without touching page order; +update-slide overwrites an entire page — anything missing from --content is deleted. Images must be uploaded Drive file_tokens; raw http(s) URLs render as broken images because the slides renderer never proxies external links, and media tops out at 20MB with no chunking. History revert accepts history_version_id, never revision_id, and wiki links (/wiki/TOKEN) resolve to an obj_token that must pass an obj_type == "slides" check. Two routing quirks stand out: a doubao.com /slides/ URL is handled by this skill — routing keys on the path pattern and token, not the domain — and gradients must use rgba() with stops, because rgb() or missing stops silently fall back to white. Emoji are banned on every slide.

Whiteboard Brings Its Own npm Package: lark-whiteboard#

lark-whiteboard (v1.0.0, 3.1KB) is the only lark skill with a second runtime dependency: before anything, it makes you verify npx -y @larksuite/whiteboard-cli@^0.2.13 -v alongside lark-cli. The decision table splits read-only from write first. +export --output-type preview|svg|source pulls rendered images, vectors, or the board’s Mermaid/PlantUML source without mutating anything. Writes go through +update with --input_format taking a single value — mermaid, plantuml, or svg — and overwriting a non-empty existing board requires an explicit “this rebuilds the whole board” confirmation. Editing an existing board with SVG walks routes/svg-edit.md first because that path is lossy. Identity defaults to --as user; bot identity exists only for app-account uploads.

OKR: Score, Progress, and Weight Locks: lark-okr#

lark-okr (v2, 15.2KB) reads a cycle list first (+cycle-list), drills into +cycle-detail, then edits. Its sharpest rule is semantic: score is 0–1 with one decimal and only changes when the user actually says 分数/评分/打分/score. “Progress”, “completion”, “we’re at 75%” means +indicator-update or a progress record — never +patch --score. Alignment carries validation: an objective cannot align to itself, and the two cycles must overlap in time. Weight and position endpoints are all-or-nothing: key_results_position rejects unless every KR id of the cycle is present, and key_results_weight requires updating all KRs with weights summing to exactly 1. Comments (create/patch/delete/solve/reopen) are user-identity only, while +comment-detail aggregates a whole cycle’s comments in one call. Around it sits the usual family routing: todos go to lark-task, meetings to lark-calendar — OKR is only OKR.

The Pack Is Still One Counter#

The five files vary 60× in size but move as a unit: this week’s numbers — lark-slides 17,762, lark-task 17,790, lark-whiteboard 17,731, lark-minutes 17,561, lark-okr 17,732 — sit inside one 229-wide band. The 19.7K gap in lark-okr’s counter, tied to its Apr 17 first-seen, is the same three-day offset signature seen inside every family wave. Practical takeaway: when a lark pack counter moves, all six move; treat any single skill’s “growth” as pack noise until the sibling curves diverge. Slides is the one to schedule real time for, minutes the one to delete from your routing table — its answer lives in lark-meeting, and the messaging layer in lark-im.