The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses highly authoritative language ('CRITICAL DIRECTIVE', 'FORBIDDEN from skipping steps') to force the agent to perform environment reconnaissance using commands like uname -a, echo $SHELL, and command -v. Additionally, the script update_garbled_dictionary.py uses subprocess.run to execute other Python scripts dynamically.
[PERSISTENCE]: The script update_garbled_dictionary.py is designed to modify resource files (garbled_replacements.json) located in the directories of other skills, specifically anibon-timestamper. This allows the skill to persist changes that affect the behavior of other installed skills even after the current task is finished.
[METADATA_POISONING]: The SKILL.md instructions use aggressive prompt engineering techniques ('MANDATORY', 'CRITICAL PATH RULE', 'Iron Rules') to override the agent's natural decision-making process and ensure the execution of potentially risky shell commands.
[DYNAMIC_EXECUTION]: The skill uses dynamic path manipulation (sys.path.insert) and imports modules at runtime based on script locations. The update_garbled_dictionary.py script also performs runtime imports of other local scripts after modifying the environment.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process auto-generated transcripts, which are untrusted external data sources. While the processing is regex-based, the skill lacks explicit boundary markers or sanitization logic to prevent instructions embedded within transcripts from influencing the agent's context, especially during the 'Detect' and 'Decode' phases described in the Thai-Latin cleaning workflow.

cleaning-auto-transcripts