The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes session data from git log and git diff to extract patterns. This untrusted data could theoretically contain instructions intended to influence the sub-agents. The skill mitigates this through a four-stage human-in-the-loop review process (Phase 1 selection, Phase 2 similarity decision, Phase 3 blueprint approval, and Phase 5 registration).
[DYNAMIC_EXECUTION]: As part of its core functionality, the skill generates new executable scripts and instructions. It includes a validation step where the sub-agent duri is instructed to execute these newly created scripts (e.g., using --help or --dry-run) to ensure they are functional before registration.
[COMMAND_EXECUTION]: The skill executes local shell commands to scan the existing skill inventory and analyze the current git workspace. These operations are restricted to the user's project directory and the standard Claude Code configuration directory (~/.claude).

skill-factory