The Agent Skills Directory

[SAFE]: The skill operates entirely on Markdown files and standard Git operations. It does not introduce any third-party dependencies, remote script executions, or network-based exfiltration vectors.
[DATA_EXPOSURE]: The instructions emphasize 'source grounding' and treat files in raw/ as immutable. This creates a safe boundary where the AI reads sensitive or raw data to synthesize information but is explicitly forbidden from altering the source of truth.
[COMMAND_EXECUTION]: While the skill mentions using Git, it does so within the context of version control for text files. There are no patterns suggesting arbitrary shell execution or privilege escalation.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest external data (sources in raw/). It acknowledges the risk of 'source-specific disagreement' and 'contradictions,' instructing the agent to explicitly distinguish source facts from new analysis and preserve provenance. This serves as a natural defensive boundary against malicious instructions embedded in ingested documents.
[REMOTE_CODE_EXECUTION]: The skill mentions optional CLI tools like qmd or Marp as suggestions, but does not provide scripts that automatically download or execute them, leaving tool adoption to the user's discretion.

llm-wiki