The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes records and data from Lark Base, which constitutes untrusted external input that could potentially contain malicious instructions.
Ingestion points: Untrusted data enters the agent context through +record-list, +record-get, and +record-search commands.
Boundary markers: The instructions mandate the use of structured NDJSON artifacts for data storage and processing, which helps separate data from control.
Capability inventory: The skill uses lark-cli for network operations and file writes, and provides templates for python3 and jq to process the resulting files locally.
Sanitization: Processing relies on JSON parsing (json.loads in Python, jq), which reduces the risk of direct command injection compared to string interpolation in shells.
[DYNAMIC_EXECUTION]: The skill provides Python script templates to be executed locally via a heredoc to process local data files.
Evidence: SKILL.md contains a template for `python3
<<'PY'` to summarize record data.
Context: The scripts are based on static, known templates provided within the skill and operate exclusively on local data artifacts.
[EXTERNAL_DOWNLOADS]: The skill can fetch and copy base configurations from the official Lark Template Center.
Source: lark-base-template-center.md describes using +base-copy with a template token to create new Base resources from the official public library.

lark-base