The Agent Skills Directory

[SAFE]: The skill performs local file system operations (reading markdown files and checking timestamps) within the standard .gemini/skills directories. It does not perform network operations or access sensitive credentials.
[COMMAND_EXECUTION]: The skill executes local shell scripts (quick-diff.sh, save-results.sh, scan.sh) to manage its internal logic. These scripts use standard tools like jq, find, and awk to process local skill metadata and usage logs. Input validation and defense-in-depth measures (e.g., checking path suffixes) are implemented within the scripts to prevent accidental execution outside of the intended directories.
[DATA_EXPOSURE]: The skill reads from a local observations.jsonl file to calculate usage frequency. This data remains local and is used solely to generate the quality audit report for the user.
[INDIRECT_PROMPT_INJECTION]: The skill processes skill files as data for evaluation by a subagent. It implements boundary markers (e.g., [INVENTORY] and [CHECKLIST] labels) in the prompt to reduce the risk of instructions inside the audited skills influencing the auditor agent's behavior.

skill-stocktake