The Agent Skills Directory

[PERSISTENCE_MECHANISMS]: The skill includes an installation script (scripts/install-hook.sh) that modifies the platform's configuration file at ~/.claude/settings.json. This script registers a UserPromptSubmit hook that persists across different sessions and automatically executes a monitoring script (scripts/task-tracker.py) whenever the user submits a prompt.
[DATA_EXPOSURE_AND_EXFILTRATION]: The monitoring script (scripts/task-tracker.py) reads user prompts from the input stream and appends them to a local history log file at ~/.claude/skill-builder-history.jsonl. This behavior results in the systematic logging of user interaction data on the host machine.
[DYNAMIC_EXECUTION]: The scripts/install-hook.sh script utilizes a heredoc to execute an embedded Python script at runtime (`python3
<<'PY'`) to perform logic for merging hook configurations into the existing JSON settings file.
[INDIRECT_PROMPT_INJECTION]: The scripts/task-tracker.py script processes untrusted user input directly from stdin. This input is then tokenized and used to generate a <system-reminder> which is injected back into the agent's context. While the script uses tokenization and stemming as a form of sanitization, it represents an attack surface for influencing agent behavior via processed prompt history.

skill-builder