The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes a large volume of untrusted markdown data from personal notes and meeting transcripts while possessing the capability to write to the filesystem and push to remote repositories. This creates a surface where embedded instructions in the notes could potentially influence the agent's behavior during the sanitization process.
Ingestion points: Full repository scan across multiple directories including people/, meetings/, daily/, projects/, and analysis/.
Boundary markers: The skill lacks explicit prompt-level delimiters (such as XML tags or unique markers) to isolate the data being sanitized from the system instructions.
Capability inventory: File system read/write access, shell command execution (git, rsync, grep), and network transmission via git push.
Sanitization: Instructions guide the model to remove sensitive content based on judgment, but do not provide specific escaping protocols for the content being ingested into the LLM context.
[COMMAND_EXECUTION]: The skill instructs the agent to execute powerful and potentially destructive shell commands to manage repository history and synchronization.
Evidence: Use of git filter-repo to rewrite repository history and git push --force to update remote branches.
Mitigation: The skill incorporates a "data-loss-gate" confirmation protocol, requiring the user to explicitly type "yes" or "do it" before executing destructive history purges or bulk edits.
[DATA_EXFILTRATION]: The skill is designed to explicitly read and process highly sensitive personal and corporate data, which is then transmitted to a shared repository.
Evidence: The scan scope includes compensation (salary, equity), performance assessments, and internal management dynamics. While the purpose is to sanitize this data, the workflow involves reading these sensitive files and subsequently pushing a modified version to a user-specified remote repository URL.

company-brainify