The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/align_all.py script coordinates multi-step alignment tasks by executing sibling Python scripts.
Evidence: The script uses subprocess.run(cmd, capture_output=True, text=True, timeout=30) to call smart_stitch.py. The command is passed as a list, which is a secure practice that prevents shell injection.
[INDIRECT_PROMPT_INJECTION]: The skill's core functionality creates a potential feedback loop where untrusted data from a user conversation is persisted and then injected into system instructions.
Ingestion points: Raw user feedback from the dialogue is extracted by the agent, stored in evolution.json via scripts/merge_evolution.py, and subsequently written into the SKILL.md file by scripts/smart_stitch.py.
Boundary markers: The scripts/smart_stitch.py script attempts to delineate the injected content by prepending a warning note (> **Auto-Generated Section**...). While helpful for human readers, this may not prevent a language model from adhering to malicious instructions embedded within the auto-generated section.
Capability inventory: The skill possesses file-write capabilities for SKILL.md and evolution.json, as well as the ability to trigger local script execution via subprocess.run.
Sanitization: The implementation performs standard JSON validation and string stripping but lacks specific filtering or sanitization to detect and remove prompt injection markers (e.g., commands intended to override agent safety guidelines) from the user-provided experience data.

skill-evolution