The Agent Skills Directory

[SAFE]: The skill is entirely composed of Markdown-based instructional content, workflows, and code recipes. No executable scripts or automation are included.
[PROMPT_INJECTION]: Detections of 'ignore previous instructions' in references/improve.md and references/opportunities.md are false positives. In context, these are security instructions to the agent (Hard Rule 4) to treat analyzed repository content as inert data and explicitly ignore any injection attempts found within that data.
[INDIRECT_PROMPT_INJECTION]: The skill includes instructions to process external codebases, creating a potential attack surface.
Ingestion points: Reads repository files during the 'Audit' phase (references/improve.md).
Boundary markers: Explicitly instructs the agent to treat file contents as 'inert' and 'data, not instructions'.
Capability inventory: Uses Read and Write tools to analyze code and generate plan files under plans/.
Sanitization: Instructs the agent to flag and ignore any steering attempts found in analyzed files.
[DATA_EXPOSURE_EXFILTRATION]: No network operations or access to sensitive local environment files (e.g., .env, SSH keys) are present in the instructions.
[REMOTE_CODE_EXECUTION]: The skill does not perform any external package installations or execute code from remote sources.

kernel-animate