The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a broad attack surface for indirect prompt injection because it is designed to ingest and process data from Lark documents, messages, and records.
Ingestion points: Data enters the agent's context through tools like docs +fetch (Lark documents) and base +record-list (Bitable records).
Boundary markers: While the skill uses structured XML and NDJSON for data handling, it lacks consistent 'ignore embedded instructions' warnings when processing retrieved content.
Capability inventory: The agent has extensive capabilities to modify the environment, including editing documents, executing SQL on application databases, and sending messages to users and groups.
Sanitization: The skill relies on platform-level security and explicit user confirmation gates (exit code 10 protocol) rather than automated content sanitization.
[EXTERNAL_DOWNLOADS]: The skill fetches required runtime assets and tools from well-known and trusted service domains associated with the Lark/Feishu platform.
Evidence: Generated HTML templates download React, Babel, and ECharts libraries from ByteDance's official CDN (sf3-scmcdn-cn.feishucdn.com).
CLI Tools: The skill guides the user to install the official @larksuite/cli package from the npm registry.
[COMMAND_EXECUTION]: The skill utilizes the lark-cli binary to perform authenticated operations on the Lark platform. High-impact operations such as SQL execution (+db-execute), application deployment (+release-create), and cache clearing (+cache-clear) are protected by a 'confirmation gate' requiring an explicit --yes flag or user approval.
[DYNAMIC_EXECUTION]: The 'Creative Design' and 'Interactive Prototype' modules generate and execute HTML, CSS, and JSX code within the browser. This is an intended primary feature for providing interactive mocks and presentations, with code using React and transpiled at runtime via Babel.

lark