The Agent Skills Directory

[DATA_EXPOSURE]: The skill is designed to read highly sensitive local files located at ~/.cursor/projects/<slug>/agent-transcripts/<uuid>/<uuid>.jsonl. These transcripts contain the full history of user-agent interactions, which likely includes proprietary source code, internal system paths, and potentially hardcoded credentials or personal data shared during previous sessions.
[COMMAND_EXECUTION]: The skill uses local shell commands including ls -t to determine file modification times and grep to search through transcript contents. It also executes git and gh (GitHub CLI) to verify the status of branches, pull requests, and tickets.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from a 'shared record' comprising external sources like issue trackers, user reports, and public chat channels via the why skill. This creates a vulnerability surface where an external attacker could embed malicious instructions in a bug report or chat message that the agent might obey when reconstructing the working context.
[DYNAMIC_EXECUTION]: The skill parallelizes its workload by spawning subagents using a 'fast, cheap model' to process slices of the transcript corpus. This dynamic task delegation involves passing processed snippets of local data to secondary agent instances.

recall