The Agent Skills Directory

[SAFE]: The skill defines procedures for corpus-level maintenance using the skill-graph toolset. Its behavior is consistent with its stated purpose of systematically auditing and improving project-specific agent skills.
[COMMAND_EXECUTION]: The skill teaches the agent how to execute the skill-graph evolve command with various flags (e.g., --auto-improve, --continuous). This is the primary intended function of the skill within its workspace.
[EXTERNAL_DOWNLOADS]: The skill references its official project repository (github.com/jacob-balslev/skill-graph) for grounding and integrity verification. This is a trusted resource provided by the author for project functionality.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill uses a sidecar file (audit-state.json) to track the audit status and integrity of the skill library. This is a recognized best practice for maintaining a verifiable chain of trust in agent skill management.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface as it is designed to ingest and process a library of external skills.
Ingestion points: Skill library corpus (configured via --skills-dir).
Boundary markers: Not explicitly defined in instructions for processed content.
Capability inventory: File modification (improve operation), CLI execution (evolve), and script evaluation (evaluate runner).
Sanitization: Not mentioned in the instructional text. This surface is inherent to the tool's primary purpose of corpus management and does not indicate malicious intent.

skill-evolution