The Agent Skills Directory

Persistence Mechanisms (HIGH): The skill creates permanent hooks in .claude/settings.json and executable shell scripts in .claude/hooks/. These are configured to run automatically on system events like SessionEnd, creating a persistent execution vector for generated code.
Dynamic Execution (HIGH): The skill writes shell scripts (.sh) to the filesystem and uses chmod +x to grant execution permissions. It also defines a workflow for compiling TypeScript into JavaScript and executing it via node, effectively allowing the agent to write and run its own software.
Indirect Prompt Injection (LOW):
Ingestion points: Reads untrusted session history and 'learnings' files from .claude/cache/learnings/*.md.
Boundary markers: Absent. The skill extracts content from markdown headers and table sections without using delimiters or instructions to ignore embedded commands.
Capability inventory: The skill possesses the Bash, Write, and Edit tools, allowing it to modify core configuration files and create new executable scripts.
Sanitization: Absent. There is no validation or sanitization of the 'learned' patterns before they are written into system-level rules or executable hooks.
Command Execution (MEDIUM): Uses Bash to perform filesystem management and script creation. The use of chmod on dynamically generated content is a high-risk pattern.

compound-learnings