The Agent Skills Directory

[DATA_EXPOSURE]: The skill is designed to read agent transcripts stored in the local file system at ~/.cursor/projects/<slug>/agent-transcripts/. These files contain historical chat logs which are inherently sensitive as they may contain previously discussed code, logic, or project details.
[COMMAND_EXECUTION]: The instructions direct the agent to utilize standard command-line tools such as ls -t for sorting files by modification time, grep for searching patterns within logs, and git/gh for synchronizing state with version control and issue trackers.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple sources including past transcripts, issue trackers, and chat channels, which could contain instructions designed to influence the agent's behavior.
Ingestion points: Agent chat transcripts (.jsonl files), source control systems, issue trackers, and error tracking logs.
Boundary markers: The instructions lack explicit delimiters for incoming data but include a mandate to "sanitize private context" before generating public output.
Capability inventory: The agent has capabilities to read files, execute local search commands, and interact with repository management tools (Git/GitHub).
Sanitization: The skill relies on an internal prompt instruction to filter sensitive data before final delivery.

recall