The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements an indirect prompt injection feedback loop by design, which can be exploited to modify agent behavior through untrusted data.
Ingestion points: The agent is instructed to capture content from tool outputs via the CLAUDE_TOOL_OUTPUT environment variable and user chat corrections into markdown files in the .learnings/ directory.
Boundary markers: The captured data is appended to log files without boundary markers or instructions to disregard embedded commands, making it susceptible to injection.
Capability inventory: The skill encourages the agent to promote logged content to high-priority instruction files such as CLAUDE.md, AGENTS.md, and SOUL.md. It also provides a script (extract-skill.sh) to generate new executable skills.
Sanitization: There is no automated or instructed sanitization process to filter or validate external content before it is integrated into the agent's long-term memory or instructions.
[COMMAND_EXECUTION]: The skill relies on local shell scripts (scripts/activator.sh, scripts/error-detector.sh) configured as hooks that run automatically during the agent's session (e.g., on prompt submission or tool output). It also provides scripts/extract-skill.sh which performs file and directory operations based on user-provided skill names.
[EXTERNAL_DOWNLOADS]: The skill documentation suggests installation from external sources, including a GitHub repository and the ClawdHub package registry.

self-improvement