The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content from multiple ingestion points, including Markdown and HTML files from DingDrive, and data from external API responses. It implements a defensive strategy by explicitly instructing the agent to treat these contents as inert data and ignore any embedded natural language instructions. Documentation in references/html.md and references/markdown.md specifically warns against executing content found within these files.
[DYNAMIC_EXECUTION]: To support the Yida low-code platform, the skill contains logic for converting JSX code into valid component schemas through a dedicated transformation pipeline. While this involves code transformation, it is restricted to the skill's primary purpose of application development and is mitigated by built-in linting checks and a focus on generating conservative, runtime-compatible code.
[COMMAND_EXECUTION]: Several automation scripts execute system commands using the platform's standard CLI tool. These operations are strictly governed by workflow references that require pre-flight checks, dry-runs, and explicit user consent before any state-changing actions are performed, including path-safety checks for local file access.
[EXTERNAL_DOWNLOADS]: The skill facilitates the discovery of official API contracts by fetching documentation from well-known service domains and downloading user-specified media (like attendance photos) for report generation. These network operations target trusted infrastructure and are performed in the context of user-initiated tasks.

dingtalk-misc